SimGrid
3.21
Versatile Simulation of Distributed Systems
|
As explained in the introduction, any SimGrid study must entail the description of the platform on which you want to simulate your application. You have to describe each element of your platform, such as computing hosts, clusters, each disks, links, etc. You must also define the routing on your platform, ie which path is taken between two hosts. Finally, you may also describe an experimental scenario, with qualitative changes (e.g., bandwidth changes representing an external load) and qualitative changes (representing how some elements fail and restart over time).
You should really separate your application from the platform description, as it will ease your experimental campain afterward. Mixing them is seen as a really bad experimental practice. The easiest to enforce this split is to put the platform description in a XML file. Many example platforms are provided in the archive, and this page gives all needed details to write such files, as well as some hints and tricks about describing your platform.
On the other side, XML is sometimes not expressive enough for some platforms, in particular large platforms exhibiting repetitive patterns that are not simply expressed in XML. In practice, many users end up generating their XML platform files from some sort of scripts. It is probably preferable to rewrite your XML platform using the lua scripting language instead. In the future, it should be possible to describe the platform directly in C++, but this is not possible yet.
As usual, SimGrid is a versatile framework, and you should find the way of describing your platform that best fits your experimental practice.
Your platform description should follow the specification presented in the simgrid.dtd DTD file. The same DTD is used for both the platform and deployment files.
From time to time, this DTD evolves to introduce possibly backward-incompatible changes. That is why each platform desciption is enclosed within a platform
tag, that have a version
attribute. The current version is 4.1. The simgrid_update_xml
program can upgrade most of the past platform files to the recent formalism.
Here is a very simple platform file, containing 3 resources (two hosts and one link), and explicitly giving the route between the hosts.
As we said, the englobing <platform> tag is used to specify the dtd version used for this file.
Then, every resource (specified with <host>, <link> or others) must be located within a given networking zone. Each zone is in charge of the routing between its resources. It means that when an host wants to communicate with another host of the same zone, it is the zone's duty to find the list of links that are involved in the communication. Here, since the <zone> tag has Full as a routing attribute, all routes must be explicitely given using the <route> and <link_ctn> tags (this routing model is both simple and inefficient :) It is OK to not specify the route between two hosts, as long as the processes located on them never try to communicate together.
A zone can contain several zones itself, leading to a hierarchical decomposition of the platform. This can be more efficient (as the inter-zone routing gets factorized with <zoneRoute>), and allows to have more than one routing model in your platform. For example, you could have a coordinate-based routing for the WAN parts of your platforms, a full routing within each datacenter, and a highly optimized routing within each cluster of the datacenter. In this case, determining the route between two given hosts gets somewhat more complex but SimGrid still computes these routes for you in a time- and space-efficient manner. Here is an illustration of these concepts:
Circles represent processing units and squares represent network routers. Bold lines represent communication links. The zone "AS2" models the core of a national network interconnecting a small flat cluster (AS4) and a larger hierarchical cluster (AS5), a subset of a LAN (AS6), and a set of peers scattered around the world (AS7).
An host is the computing resource on which an actor can execute.
Attribute | Values | Description |
---|---|---|
id | String (mandatory) | The identifier of the host. facilitates referring to this AS. |
speed | double (mandatory) | Computational power of every core of this host in FLOPS (must be positive) |
core | int (defaults to 1) | Number of cores (see How to model multicore machines) |
availability_file | File name (optional) | (Relative or absolute) filename to use as input; must contain availability traces for this host. The syntax of this file is defined below. |
state_file | File name (optional) | File to use as a state profile (see How to model churn) |
coordinates | String (mandatory when using Vivaldi routing) | The coordinates of this host (see P2P or how to use coordinates). |
pstate | Double (Defaults to 0) | FIXME: Not yet documented. |
SimGrid provides mechanisms to change a hosts' availability over time, using the availability_file
attribute to the @<host@>
tag and a separate text file whose syntax is exemplified below.
<platform version="4"> <host id="bob" speed="500Gf" availability_file="bob.trace" /> </platform>
Let us begin to explain this example by looking at line 2. (Line 1 will become clear soon). The first column describes points in time, in this case, time 0. The second column describes the relative amount of power this host is able to deliver (relative to the maximum performance specified in the @<host@>
tag). (Clearly, the second column needs to contain values that are not smaller than 0 and not larger than 1). In this example, our host will deliver 500 Mflop/s at time 0, as 500 Mflop/s is the maximum performance of this host. At time 11.0, it will deliver half of its maximum performance, i.e., 250 Mflop/s until time 20.0 when it will will start delivering 80% of its power. In this example, this amounts to 400 Mflop/s.
Since the periodicity in line 1 was set to be 1.0, i.e., 1 timestep, this host will continue to provide 500 Mflop/s from time 21. From time 32 it will provide 250 MFlop/s and so on.
<cluster />
represents a machine-cluster. It is most commonly used when one wants to define many hosts and a network quickly. Technically, cluster
is a meta-tag: from the inner SimGrid point of view, a cluster is a network zone where some optimized routing is defined. The default inner organization of the cluster is as follow:
__________ | | | router | ____________|__________|_____________ backbone | | | | | | l0| l1| l2| l97| l96 | | l99 | | | ........ | | | | | c-0.me c-99.me
Here, a set of hosts is defined. Each of them has a link to a central backbone (backbone is a link itself, as a link can be used to represent a switch, see the switch / link section below for more details about it). A router allows to connect a cluster to the outside world. Internally, SimGrid treats a cluster as a network zone containing all hosts: the router is the default gateway for the cluster.
There is an alternative organization, which is as follows:
__________ | | | router | |__________| / | @ / | @ l0 / l1| @l2 / | @ / | @ host0 host1 host2
The principle is the same, except that there is no backbone. This representation can be obtained easily: just do not set the bb_* attributes.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | The identifier of the cluster. Facilitates referring to this cluster. |
prefix | yes | string | Each node of the cluster has to have a name. This name will be prefixed with this prefix. |
suffix | yes | string | Each node of the cluster will be suffixed with this suffix |
radical | yes | string | Regexp used to generate cluster nodes name. Syntax: "10-20" will give you 11 machines numbered from 10 to 20, "10-20;2" will give you 12 machines, one with the number 2, others numbered as before. The produced number is concatenated between prefix and suffix to form machine names. |
speed | yes | int | Same as the speed attribute of the @<host@> tag. |
core | no | int (default: 1) | Same as the core attribute of the @<host@> tag. |
bw | yes | int | Bandwidth for the links between nodes and backbone (if any). See the link section for syntax/details. |
lat | yes | int | Latency for the links between nodes and backbone (if any). See link section for syntax/details. |
sharing_policy | no | string | Sharing policy for the links between nodes and backbone (if any). See link section for syntax/details. |
bb_bw | no | int | Bandwidth for backbone (if any). See link section for syntax/details. If bb_bw and bb_lat (see below) attributes are omitted, no backbone is created (alternative cluster architecture described before). |
bb_lat | no | int | Latency for backbone (if any). See link section for syntax/details. If bb_lat and bb_bw (see above) attributes are omitted, no backbone is created (alternative cluster architecture described before). |
bb_sharing_policy | no | string | Sharing policy for the backbone (if any). See link section for syntax/details. |
limiter_link | no | int | Bandwidth for limiter link (if any). This adds a specific link for each node, to set the maximum bandwidth reached when communicating in both directions at the same time. In theory this value should be 2*bw for splitduplex links, but in reality this might be less. This value will depend heavily on the communication model, and on the cluster's hardware, so no default value can be set, this has to be measured. More details can be obtained in "Toward Better Simulation of MPI Applications on Ethernet/TCP Networks" |
loopback_bw | no | int | Bandwidth for loopback (if any). See link section for syntax/details. If loopback_bw and loopback_lat (see below) attributes are omitted, no loopback link is created and all intra-node communication will use the main network link of the node. Loopback link is a FATPIPE . |
loopback_lat | no | int | Latency for loopback (if any). See link section for syntax/details. See loopback_bw for more info. |
topology | no | FLAT|TORUS|FAT_TREE|DRAGONFLY (default: FLAT) | Network topology to use. SimGrid currently supports FLAT (with or without backbone, as described before), TORUS , FAT_TREE, and DRAGONFLY attributes for this tag. topo_parameters | no | string | Specific parameters to pass for the topology defined in the topology tag. For torus networks, comma-separated list of the number of nodes in each dimension of the torus. Please refer to the specific documentation for FatTree NetZone, Dragonfly NetZone.
the router name is defined as the resulting String in the following java line of code:
router_name = prefix + clusterId + "_router" + suffix;
Consider the following two (and independent) uses of the cluster
tag:
<cluster id="my_cluster_1" prefix="" suffix="" radical="0-262144" speed="1e9" bw="125e6" lat="5E-5"/> <cluster id="my_cluster_2" prefix="c-" suffix=".me" radical="0-99" speed="1e9" bw="125e6" lat="5E-5" bb_bw="2.25e9" bb_lat="5E-4"/>
The second example creates one router and 100 machines with the following names:
c-my_cluster_2_router.me c-0.me c-1.me c-2.me ... c-99.me
Cluster
.The <cabinet />
tag is, like the <cluster> tag, a meta-tag. This means that it is simply a shortcut for creating a set of (homogenous) hosts and links quickly; unsurprisingly, this tag was introduced to setup cabinets in data centers quickly. Unlike <cluster>, however, the <cabinet> assumes that you create the backbone and routers yourself; see our examples below.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | The identifier of the cabinet. Facilitates referring to this cluster. |
prefix | yes | string | Each node of the cabinet has to have a name. This name will be prefixed with this prefix. |
suffix | yes | string | Each node of the cabinet will be suffixed with this suffix |
radical | yes | string | Regexp used to generate cabinet nodes name. Syntax: "10-20" will give you 11 machines numbered from 10 to 20, "10-20;2" will give you 12 machines, one with the number 2, others numbered as before. The produced number is concatenated between prefix and suffix to form machine names. |
speed | yes | int | Same as the speed attribute of the <host> tag. |
bw | yes | int | Bandwidth for the links between nodes and backbone (if any). See the link section for syntax/details. |
lat | yes | int | Latency for the links between nodes and backbone (if any). See the link section for syntax/details. |
The following example was taken from examples/platforms/meta_cluster.xml
and shows how to use the cabinet tag.
<zone id="my_cluster1" routing="Cluster"> <cabinet id="cabinet1" prefix="host-" suffix=".cluster1" speed="1Gf" bw="125MBps" lat="100us" radical="1-10"/> <cabinet id="cabinet2" prefix="host-" suffix=".cluster1" speed="1Gf" bw="125MBps" lat="100us" radical="11-20"/> <cabinet id="cabinet3" prefix="host-" suffix=".cluster1" speed="1Gf" bw="125MBps" lat="100us" radical="21-30"/> <backbone id="backbone1" bandwidth="2.25GBps" latency="500us"/> </zone>
The hosts generated in the above example are named host-1.cluster, host-2.cluster1 etc.
This tag represents a peer, as in Peer-to-Peer (P2P) networks. This can only be used in Vivaldi NetZones. It creates the following resources to the NetZone:
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | The identifier of the peer. Facilitates referring to this peer. |
speed | yes | int | See the description of the host tag for this attribute |
bw_in | yes | int | Bandwidth of the private downstream link |
bw_out | yes | int | Bandwidth of the private upstream link |
coordinates | no | string | Coordinates of the gateway for this peer. Example value: 12.8 14.4 6.4 |
sharing_policy | no | SHARED|SPLITDUPLEX (default: SPLITDUPLEX) | Sharing policy for links. See link description for details. availability_file| no | string | Availability file for the peer. Same as host availability file. See host description for details. state_file | no | string | State file for the peer. Same as host state file. See host description for details.
The communication latency between an host A=(xA,yA,zA) and an host B=(xB,yB,zB) is computed as follows:
latency = sqrt( (xA-xB)² + (yA-yB)² ) + zA + zB
See the documentation of simgrid::kernel::routing::VivaldiZone for details on how the latency is computed from the coordinate, and on the the up and down bandwidth are used.
There are two tags at all times available to represent network entities and several other tags that are available only in certain contexts.
<link>
: Represents a entity that has a limited bandwidth, a latency, and that can be shared according to TCP way to share this bandwidth. <router/>
: Represents an entity that a message can be routed to, but that is unable to execute any code. In SimGrid, routers have also no impact on the performance: Routers do not limit any bandwidth nor do they increase latency. As a matter of fact, routers are (almost) ignored by the simulator when the simulation has begun.<backbone/>
: This tag is only available when the containing network zone is used as a cluster (i.e., mode="Cluster")<link>
(see section). Routers are used to run some routing algorithm and determine routes (see Section Routing for details).As said before, router is used only to give some information for routing algorithms. So, it does not have any attributes except:
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | The identifier of the router to be used when referring to it. |
coordinates | no | string | Must be provided when choosing the Vivaldi, coordinate-based routing model for the network zone the router belongs to. More details can be found in the Section P2P or how to use coordinates. |
<router id="gw_dc1_horizdist"/>
Network links can represent one-hop network connections. They are characterized by their id and their bandwidth; links can (but may not) be subject to latency.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | The identifier of the link to be used when referring to it. |
bandwidth | yes | string | Maximum bandwidth for this link, along with its unit. |
latency | no | double (default: 0.0) | Latency for this link. |
sharing_policy | no | SHARED|FATPIPE|SPLITDUPLEX (default: SHARED) | Sharing policy for the link. bandwidth_file | no | string | Allows you to use a file as input for bandwidth. latency_file | no | string | Allows you to use a file as input for latency. state_file | no | string | Allows you to use a file as input for states.
latency
When using the latency attribute, you can specify the latency by using the scientific notation or by using common abbreviations. For instance, the following three tags are equivalent:
<link id="LINK1" bandwidth="125000000" latency="5E-6"/> <link id="LINK1" bandwidth="125000000" latency="5us"/> <link id="LINK1" bandwidth="125000000" latency="0.000005"/>
Here, the second tag uses "us", meaning "microseconds". Other shortcuts are:
Name | Abbreviation | Time (in seconds) |
---|---|---|
Week | w | 7 * 24 * 60 * 60 |
Day | d | 24 * 60 * 60 |
Hour | h | 60 * 60 |
Minute | m | 60 |
Second | s | 1 |
Millisecond | ms | 0.001 = 10^(-3) |
Microsecond | us | 0.000001 = 10^(-6) |
Nanosecond | ns | 0.000000001 = 10^(-9) |
Picosecond | ps | 0.000000000001 = 10^(-12) |
By default a network link is SHARED, i.e., if two or more data flows go through a link, the bandwidth is shared fairly among all data flows. This is similar to the sharing policy TCP uses.
On the other hand, if a link is defined as a FATPIPE, each flow going through this link will be provided with the complete bandwidth, i.e., no sharing occurs and the bandwidth is only limiting each flow individually. Please note that this is really on a per-flow basis, not only on a per-host basis! The complete bandwidth provided by this link in this mode is number_of_flows*bandwidth
, with at most bandwidth
being available per flow.
Using the FATPIPE mode allows to model backbones that won't affect performance (except latency).
The last mode available is SPLITDUPLEX. This means that SimGrid will automatically generate two links (one carrying the suffix _UP and the other the suffix _DOWN) for each <link>
tag. This models situations when the direction of traffic is important.
In other words: The SHARED policy defines a physical limit for the bandwidth. The FATPIPE mode defines a limit for each application, with no upper total limit.
<link id="SWITCH" bandwidth="125000000" latency="5E-5" sharing_policy="FATPIPE" />
Similar to hosts, it is possible to declare links whose state, bandwidth or latency changes over time (see Section pf_host_dynamism for details).
In the case of network links, the bandwidth
and latency
attributes are replaced by the bandwidth_file
and latency_file
attributes. The following XML snippet demonstrates how to use this feature in the platform file. The structure of the files "link1.bw" and "link1.lat" is shown below.
<link id="LINK1" state_file="link1.fail" bandwidth="80000000" latency=".0001" bandwidth_file="link1.bw" latency_file="link1.lat" />
In this example, the bandwidth changes repeatedly, with all changes being repeated every 12 seconds.
At the beginning of the the simulation, the link's bandwidth is 80,000,000 B/s (i.e., 80 Mb/s); this value was defined in the XML snippet above. After four seconds, it drops to 40 Mb/s (line 2), and climbs back to 60 Mb/s after another 4 seconds (line 3). The value does not change any more until the end of the period, that is, after 12 seconds have been simulated). At this point, periodicity kicks in and this behavior is repeated: Seconds 12-16 will experience 80 Mb/s, 16-20 40 Mb/s etc.).
In this example, the latency varies with a period of 5 seconds. In the xml snippet above, the latency is initialized to be 0.0001s (100µs). This value will be kept during the first second, since the latency_file contains changes to this value at second one, two and three. At second one, the value will be 0.001, i.e., 1ms. One second later it will be adjusted to 0.01 (or 10ms) and one second later it will be set again to 1ms. The value will not change until second 5, when the periodicity defined in line 1 kicks in. It then loops back, starting at 100µs (the initial value) for one second.
<prop/>
tagSimilar to the <host>
tag, a link may also contain the <prop/>
tag; see the host documentation (Section <host>) for an example.
Using this tag, you can designate an already existing link to be a backbone.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | Name of the link that is supposed to act as a backbone. |
The storage facilities implemented in SimGrid help to model (and account for) storage devices, such as tapes, hard-drives, CD or DVD devices etc. A typical situation is depicted in the figure below:
In this figure, two hosts called Bob and Alice are interconnected via a network and each host is physically attached to a disk; it is not only possible for each host to mount the disk they are attached to directly, but they can also mount disks that are in a remote location. In this example, Bob mounts Alice's disk remotely and accesses the storage via the network.
SimGrid provides 3 different entities that can be used to model setups that include storage facilities:
Entity name | Description |
---|---|
storage_type | Defines a template for a particular kind of storage (such as a hard-drive) and specifies important features of the storage, such as capacity, performance (read/write), contents, ... Different models of hard-drives use different storage_types (because the difference between an SSD and an HDD does matter), as they differ in some specifications (e.g., different sizes or read/write performance). |
storage | Defines an actual instance of a storage type (disk, RAM, ...); uses a storage_type template (see line above) so that you don't need to re-specify the same details over and over again. |
mount | Must be wrapped by a <host> tag; declares which storage(s) this host has mounted and where (i.e., the mountpoint). |
In order to assess exactly how much time is spent reading from the storage, SimGrid needs to know what is stored on the storage device (identified by distinct (file-)name, like in a file system) and what size this content has.
Here is an excerpt from two storage content file; if you want to see the whole file, check the file examples/platforms/content/storage_content.txt
that comes with the SimGrid source code.
SimGrid essentially supports two different formats: UNIX-style filepaths should follow the well known format:
/lib/libsimgrid.so.3.6.2 12710497 /bin/smpicc 918 /bin/smpirun 7292 /bin/smpif2c 1990 /bin/simgrid_update_xml 5018 /bin/graphicator 66986 /bin/simgrid-colorizer 2993 /bin/smpiff 820 /bin/tesh 356434
Windows filepaths, unsurprisingly, use the windows style:
@Windows@avastSS.scr 41664 @Windows@bfsvc.exe 75264 @Windows@bootstat.dat 67584 @Windows@CoreSingleLanguage.xml 31497 @Windows@csup.txt 12 @Windows@dchcfg64.exe 335464 @Windows@dcmdev64.exe 93288
If you want to generate a storage content file based on your own filesystem (or at least a filesystem you have access to), try running this command (works only on unix systems):
find . -type f -exec ls -1s --block=1 {} @; 2>/dev/null | awk '{ print $2 " " $1}' > ./content.txt
These are the entities that you can use in your platform files to include storage in your model. See also the list of our example files; these might also help you to get started.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | Identifier of this storage_type; used when referring to it |
model | no | string | In the future, this will allow to change the performance model to use |
size | yes | string | Specifies the amount of available storage space; you can specify storage like "500GiB" or "500GB" if you want. (TODO add a link to all the available abbreviations) |
content | yes | string | Path to a Storage Content File on your system. This file must exist. |
This tag must contain some predefined model properties, specified via the <model_prop> tag. Here is a list, see below for an example:
Property id | Mandatory | Values | Description |
---|---|---|---|
Bwrite | yes | string | Bandwidth for write access; in B/s (but you can also specify e.g. "30MBps") |
Bread | yes | string | Bandwidth for read access; in B/s (but you can also specify e.g. "30MBps") |
Here is a complete example for the storage_type
tag:
<storage_type id="single_HDD" size="4000"> <model_prop id="Bwrite" value="30MBps" /> <model_prop id="Bread" value="100MBps" /> <prop id="Brand" value="Western Digital" /> </storage_type>
Attributes | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | Identifier of this storage ; used when referring to it |
typeId | yes | string | Here you need to refer to an already existing <storage_type>; the storage entity defined by this tag will then inherit the properties defined there. |
attach | yes | string | Name of a host (see Section <host>) to which this storage is physically attached to (e.g., a hard drive in a computer) |
content | no | string | When specified, overwrites the content attribute of <storage_type> |
Here are two examples:
<storage id="Disk1" typeId="single_HDD" attach="bob" /> <storage id="Disk2" typeId="single_SSD" content="content/win_storage_content.txt" />
The first example is straightforward: A disk is defined and called "Disk1"; it is of type "single_HDD" (shown as an example of <storage_type> above) and attached to a host called "bob" (the definition of this host is omitted here).
The second storage is called "Disk2", is still of the same type as Disk1 but now specifies a new content file (so the contents will be different from Disk1) and the filesystem uses the windows style; finally, it is attached to a second host, called alice (which is again not defined here).
Attribute | Mandatory | Values | Description |
---|---|---|---|
id | yes | string | Refers to a <storage> entity that will be mounted on that computer |
name | yes | string | Path/location to/of the logical reference (mount point) of this disk |
This tag must be enclosed by a <host> tag. It then specifies where the mountpoint of a given storage device (defined by the id
attribute) is; this location is specified by the name
attribute.
Here is a simple example, taken from the file examples/platform/storage.xml
:
<storage_type id="single_SSD" size="500GiB"> <model_prop id="Bwrite" value="60MBps" /> <model_prop id="Bread" value="200MBps" /> </storage_type> <storage id="Disk2" typeId="single_SSD" content="content/win_storage_content.txt" attach="alice" /> <storage id="Disk4" typeId="single_SSD" content="content/small_content.txt" attach="denise"/> <host id="alice" speed="1Gf"> <mount storageId="Disk2" name="c:"/> </host> <host id="denise" speed="1Gf"> <mount storageId="Disk2" name="c:"/> <mount storageId="Disk4" name="/home"/> </host>
This example is quite interesting, as the same device, called "Disk2", is mounted by two hosts at the same time! Note, however, that the host called alice
is actually attached to this storage, as can be seen in the <storage> tag. This means that denise
must access this storage through the network, but SimGrid automatically takes care of that for you.
Furthermore, this example shows that denise
has mounted two storages with different filesystem types (unix and windows). In general, a host can mount as many storage devices as required.
attach
and mount
is simply that an attached storage is always physically inside (or connected to) that machine; for instance, a USB stick is attached to one and only one machine (where it's plugged-in) but it can only be mounted on others, as mounted storage can also be a remote location.examples/platforms/storage/remote_io.xml examples/platforms/storage/storage.xml
Several examples were already discussed above; if you're interested in full examples, check the the following platforms:
examples/platforms/storage.xml
examples/platforms/remote_io.xml
If you're looking for some examplary C code, you may find the source code available in the directory examples/msg/io/
useful.
The storage functionality of SimGrid is type-agnostic, that is, the implementation does not presume any type of storage, such as HDDs/SSDs, RAM, CD/DVD devices, USB sticks etc.
This allows the user to apply the simulator for a wide variety of scenarios; one common scenario would be the access of remote RAM.
How can this be achieved in SimGrid? Let's assume we have a setup where three hosts (HostA, HostB, HostC) need to access remote RAM:
Host A / RAM -- Host B @ Host C
An easy way to model this scenario is to setup and define the RAM via the storage and storage type entities and attach it to a remote dummy host; then, every host can have their own links to this host (modelling for instance certain scenarios, such as PCIe ...)
Host A / RAM - Dummy -- Host B @ Host C
Now, if read from this storage, the host that mounts this storage communicates to the dummy host which reads from RAM and sends the information back.
To achieve high performance, the routing tables used within SimGrid are static. This means that routing between two nodes is calculated once and will not change during execution. The SimGrid team chose to use this approach as it is rare to have a real deficiency of a resource; most of the time, a communication fails because the links experience too much congestion and hence, your connection stops before the timeout or because the computer designated to be the destination of that message is not responding.
We also chose to use shortest paths algorithms in order to emulate routing. Doing so is consistent with the reality: RIP, OSPF, BGP are all calculating shortest paths. They do require some time to converge, but eventually, when the routing tables have stabilized, your packets will follow the shortest paths.
Before SimGrid v3.16, networking zones used to be called Autonomous Systems, but this was misleading as zones may include other zones in a hierarchical manner. If you find any remaining reference to network zones, please report this as a bug.
Attribute | Value | Description |
---|---|---|
id | String (mandatory) | The identifier of this zone (must be unique) |
routing | One of the existing routing algorithm (mandatory) | See Section Routing models for details. |
Example:
In this example, zone0 contains two hosts (host1 and host2). The route between the hosts goes through link1.
For each network zone, you must define explicitly which routing model will be used. There are 3 different categories for routing models:
Using routers becomes mandatory when using shortest-path based models or when using the bindings to the ns-3 packet-level simulator instead of the native analytical network model implemented in SimGrid.
For graph-based shortest path algorithms, routers are mandatory, because these algorithms require a graph as input and so we need to have source and destination for each edge.
Routers are naturally an important concept ns-3 since the way routers run the packet routing algorithms is actually simulated. SimGrid's analytical models however simply aggregate the routing time with the transfer time.
So why did we incorporate routers in SimGrid? Rebuilding a graph representation only from the route information turns out to be a very difficult task, because of the missing information about how routes intersect. That is why we introduced routers, which are simply used to express these intersection points. It is important to understand that routers are only used to provide topological information.
To express this topological information, a route has to be defined in order to declare which link is connected to a router.
The following table shows all the models that compute routes using shortest-paths algorithms are currently available in SimGrid. More detail on how to choose the best routing model is given in the Section called "Choosing wisely the routing model to use".
Name | Description |
---|---|
Floyd | Floyd routing data. Pre-calculates all routes once |
Dijkstra | Dijkstra routing data. Calculates routes only when needed |
DijkstraCache | Dijkstra routing data. Handles some cache for already calculated routes. |
All those shortest-path models are instanciated in the same way and are completely interchangeable. Here are some examples:
Floyd example:
<zone id="zone0" routing="Floyd"> <cluster id="my_cluster_1" prefix="c-" suffix="" radical="0-1" speed="1000000000" bw="125000000" lat="5E-5" router_id="router1"/> <zone id="zone1" routing="None"> <host id="host1" speed="1000000000"/> </zone> <link id="link1" bandwidth="100000" latency="0.01"/> <zoneroute src="my_cluster_1" dst="zone1" gw_src="router1" gw_dst="host1"> <link_ctn id="link1"/> </zoneroute> </zone>
zoneroute given at the end gives a topological information: link1 is between router1 and host1.
This is an automatically generated list of example files that use the Floyd routing model (the path is given relative to SimGrid's source directory)
examples/platforms/bypassRoute.xml examples/platforms/cloud.xml examples/platforms/data_center.xml examples/platforms/g5k.xml
This is an automatically generated list of example files that use the Dijkstra routing model (the path is given relative to SimGrid's source directory)
examples/platforms/bypassRoute.xml
Dijkstra example:
<zone id="zone_2" routing="Dijkstra"> <host id="zone_2_host1" speed="1000000000"/> <host id="zone_2_host2" speed="1000000000"/> <host id="zone_2_host3" speed="1000000000"/> <link id="zone_2_link1" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link2" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link3" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link4" bandwidth="1250000000" latency="5E-4"/> <router id="central_router"/> <router id="zone_2_gateway"/> <!-- routes providing topological information --> <route src="central_router" dst="zone_2_host1"><link_ctn id="zone_2_link1"/></route> <route src="central_router" dst="zone_2_host2"><link_ctn id="zone_2_link2"/></route> <route src="central_router" dst="zone_2_host3"><link_ctn id="zone_2_link3"/></route> <route src="central_router" dst="zone_2_gateway"><link_ctn id="zone_2_link4"/></route> </zone>
DijkstraCache example:
<zone id="zone_2" routing="DijkstraCache"> <host id="zone_2_host1" speed="1000000000"/> ... (platform unchanged compared to upper example)
This is an automatically generated list of example files that use the DijkstraCache routing model (the path is given relative to SimGrid's source directory):
Editor's note: At the time of writing, no platform file used this routing model - so if there are no example files listed here, this is likely to be correct.
Name | Description |
---|---|
Full | You have to enter all necessary routers manually; that is, every single route. This may consume a lot of memory when the XML is parsed and might be tedious to write; i.e., this is only recommended (if at all) for small platforms. |
Full example:
<zone id="zone0" routing="Full"> <host id="host1" speed="1000000000"/> <host id="host2" speed="1000000000"/> <link id="link1" bandwidth="125000000" latency="0.000100"/> <route src="host1" dst="host2"><link_ctn id="link1"/></route> </zone>
This is an automatically generated list of example files that use the Full routing model (the path is given relative to SimGrid's source directory):
examples/msg/mc/platform.xml examples/platforms/bypassASroute.xml examples/platforms/bypassRoute.xml examples/platforms/cloud.xml examples/platforms/cluster_and_one_host.xml examples/platforms/cluster_crossbar.xml examples/platforms/cluster_dragonfly.xml examples/platforms/cluster_fat_tree.xml examples/platforms/cluster_multi.xml examples/platforms/cluster_torus.xml examples/platforms/config.xml examples/platforms/config_tracing.xml examples/platforms/crosstraffic.xml examples/platforms/data_center.xml examples/platforms/dogbone.xml examples/platforms/energy_platform.xml examples/platforms/faulty_host.xml examples/platforms/g5k.xml examples/platforms/griffon.xml examples/platforms/meta_cluster.xml examples/platforms/multicore_machine.xml examples/platforms/onelink.xml examples/platforms/prop.xml examples/platforms/routing_cluster.xml examples/platforms/simulacrum_7_hosts.xml examples/platforms/small_platform.xml examples/platforms/small_platform_fatpipe.xml examples/platforms/small_platform_one_link_routes.xml examples/platforms/small_platform_with_failures.xml examples/platforms/small_platform_with_routers.xml examples/platforms/storage/remote_io.xml examples/platforms/storage/storage.xml examples/platforms/three_multicore_hosts.xml examples/platforms/two_hosts.xml examples/platforms/two_hosts_platform_shared.xml examples/platforms/two_hosts_platform_with_availability.xml examples/platforms/two_hosts_platform_with_availability_included.xml examples/s4u/energy-boot/platform_boot.xml
Name | Description |
---|---|
Cluster | This is specific to the <cluster/> tag and should not be used by the user, as several assumptions are made. |
None | No routing at all. Unless you know what you're doing, avoid using this mode in combination with a non-constant network model. |
Vivaldi | Perfect when you want to use coordinates. Also see the corresponding P2P section below. |
This is an automatically generated list of example files that use the Cluster routing model (the path is given relative to SimGrid's source directory):
examples/platforms/meta_cluster.xml examples/platforms/routing_cluster.xml
This model does exactly what it's name advertises: Nothing. There is no routing available within this model and if you try to communicate within the zone that uses this model, SimGrid will fail unless you have explicitly activated the Constant Network Model (this model charges the same for every single communication). It should be noted, however, that you can still attach an ZoneRoute, as is demonstrated in the example below:
<?xml version='1.0'?> <!DOCTYPE platform SYSTEM "http://simgrid.gforge.inria.fr/simgrid/simgrid.dtd"> <platform version="4.1"> <zone id="AS0" routing="Full"> <cluster id="my_cluster_1" prefix="c-" suffix=".me" radical="0-1" speed="1Gf" bw="125MBps" lat="50us" router_id="router1"/> <zone id="AS1" routing="None"> <host id="host1" speed="1Gf"/> </zone> <link id="link1" bandwidth="100kBps" latency="10ms"/> <zoneRoute src="my_cluster_1" dst="AS1" gw_src="router1" gw_dst="host1"> <link_ctn id="link1"/> </zoneRoute> </zone> </platform>
This is an automatically generated list of example files that use the None routing model (the path is given relative to SimGrid's source directory):
examples/platforms/cluster_and_one_host.xml examples/platforms/prop.xml examples/platforms/routing_none.xml examples/platforms/small_platform_constant.xml
For more information on how to use the Vivaldi Coordinates, see also Section P2P tags.
Note that it is possible to combine the Vivaldi routing model with other routing models; an example can be found in the file examples/platforms/cloud.xml
. This examples models a NetZone using Vivaldi that contains other NetZones that use different routing models.
This is an automatically generated list of example files that use the None routing model (the path is given relative to SimGrid's source directory):
examples/platforms/cloud.xml examples/platforms/data_center.xml examples/platforms/two_peers.xml examples/platforms/vivaldi.xml
There are currently four different ways to define routes:
Name | Description |
---|---|
route | Used to define route between host/router |
zoneRoute | Used to define route between different zones |
bypassRoute | Used to supersede normal routes as calculated by the network model between host/router; e.g., can be used to use a route that is not the shortest path for any of the shortest-path routing models. |
bypassZoneRoute | Used in the same way as bypassRoute, but for zones |
Basically all those tags will contain an (ordered) list of references to link that compose the route you want to define.
Consider the example below:
<route src="Alice" dst="Bob"> <link_ctn id="link1"/> <link_ctn id="link2"/> <link_ctn id="link3"/> </route>
The route here from host Alice to Bob will be first link1, then link2, and finally link3. What about the reverse route? Route and zoneroute have an optional attribute symmetrical
, that can be either YES
or NO
. YES
means that the reverse route is the same route in the inverse order, and is set to YES
by default. Note that this is not the case for bypass*Route, as it is more probable that you want to bypass only one default route.
For an zoneroute, things are just slightly more complicated, as you have to give the id of the gateway which is inside the zone you want to access ... So it looks like this:
<zoneroute src="zone1" dst="zone2" gw_src="router1" gw_dst="router2"> <link_ctn id="link1"/> </zoneroute>
gw == gateway, so when any message are trying to go from zone1 to zone2, it means that it must pass through router1 to get out of the zone, then pass through link1, and get into zone2 by being received by router2. router1 must belong to zone1 and router2 must belong to zone2.
This entity has only one purpose: Refer to an already existing <link/> when defining a route, i.e., it can only occur as a child of <route/>
Attribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | String | The identifier of the link that should be added to the route. |
| direction | maybe | UP|DOWN | If the link referenced by id
has been declared as SPLITDUPLEX, this indicates which direction the route traverses through this link: UP or DOWN. If you don't use SPLITDUPLEX, do not use this attribute or SimGrid will not find the right link.
This is an automatically generated list of example files that use the <link_ctn/>
entity (the path is given relative to SimGrid's source directory):
examples/msg/mc/platform.xml examples/platforms/bypassASroute.xml examples/platforms/bypassRoute.xml examples/platforms/cloud.xml examples/platforms/cluster_and_one_host.xml examples/platforms/cluster_multi.xml examples/platforms/config.xml examples/platforms/config_tracing.xml examples/platforms/crosstraffic.xml examples/platforms/data_center.xml examples/platforms/dogbone.xml examples/platforms/energy_platform.xml examples/platforms/g5k.xml examples/platforms/griffon.xml examples/platforms/meta_cluster.xml examples/platforms/onelink.xml examples/platforms/prop.xml examples/platforms/routing_cluster.xml examples/platforms/simulacrum_7_hosts.xml examples/platforms/small_platform.xml examples/platforms/small_platform_fatpipe.xml examples/platforms/small_platform_one_link_routes.xml examples/platforms/small_platform_with_failures.xml examples/platforms/small_platform_with_routers.xml examples/platforms/storage/remote_io.xml examples/platforms/storage/storage.xml examples/platforms/three_multicore_hosts.xml examples/platforms/two_hosts.xml examples/platforms/two_hosts_platform_shared.xml examples/platforms/two_hosts_platform_with_availability.xml examples/platforms/two_hosts_platform_with_availability_included.xml examples/s4u/energy-boot/platform_boot.xml
The purpose of this entity is to define a route between two NetZones. Recall that all zones form a tree, so to connect two sibiling zones, you must give such a zoneRoute specifying the source and destination zones, along with the gateway in each zone (ie, the point to reach within that zone to reach the netzone), and the list of links in the ancestor zone to go from one zone to another.
So, to go from an host src_host
that is within zone src
, to an host dst_host
that is within dst
, you need to:
src
, from src_host
to the specified gw_src
;dst
, from gw_dst
to dst_host
.Attribute name | Mandatory | Values | Description |
---|---|---|---|
src | yes | String | The identifier of the source zone |
dst | yes | String | See the src attribute |
gw_src | yes | String | The gateway that will be used within the src zone; this can be any Host or Router defined within the src zone. |
gw_dst | yes | String | Same as gw_src , but with the dst zone instead. |
| symmetrical | no | YES|NO (Default: YES) | If this route is symmetric, the opposite route (from dst to src) will also be declared implicitly. |
<zone id="zone0" routing="Full"> <cluster id="my_cluster_1" prefix="c-" suffix=".me" radical="0-149" speed="1000000000" bw="125000000" lat="5E-5" bb_bw="2250000000" bb_lat="5E-4"/> <cluster id="my_cluster_2" prefix="c-" suffix=".me" radical="150-299" speed="1000000000" bw="125000000" lat="5E-5" bb_bw="2250000000" bb_lat="5E-4"/> <link id="backbone" bandwidth="1250000000" latency="5E-4"/> <zoneroute src="my_cluster_1" dst="my_cluster_2" gw_src="c-my_cluster_1_router.me" gw_dst="c-my_cluster_2_router.me"> <link_ctn id="backbone"/> </zoneroute> <zoneroute src="my_cluster_2" dst="my_cluster_1" gw_src="c-my_cluster_2_router.me" gw_dst="c-my_cluster_1_router.me"> <link_ctn id="backbone"/> </zoneroute> </zone>
The principle is the same as for ZoneRoute: The route contains a list of links that provide a path from src
to dst
. Here, src
and dst
can both be either a host or router. This is mostly useful for the Full routing model as well as for the shortest-paths based models (as they require topological information).
Attribute name | Mandatory | Values | Description |
---|---|---|---|
src | yes | String | The value given to the source's "id" attribute |
dst | yes | String | The value given to the destination's "id" attribute. |
| symmetrical | no | YES| NO (Default: YES) | If this route is symmetric, the opposite route (from dst to src) will also be declared implicitly. |
A route in the Full routing model could look like this:
<route src="Tremblay" dst="Bourassa"> <link_ctn id="4"/><link_ctn id="3"/><link_ctn id="2"/><link_ctn id="0"/><link_ctn id="1"/><link_ctn id="6"/><link_ctn id="7"/> </route>
A route in the Shortest-Path routing model could look like this:
<route src="Tremblay" dst="Bourassa"> <link_ctn id="3"/> </route>
As said before, once you choose a model, it (most likely; the constant network model, for example, doesn't) calculates routes for you. But maybe you want to define some of your routes, which will be specific. You may also want to bypass some routes defined in lower level zone at an upper stage: bypasszoneroute is the tag you're looking for. It allows to bypass routes defined between already defined between zone (if you want to bypass route for a specific host, you should just use byPassRoute). The principle is the same as zoneroute: bypasszoneroute contains list of links that are in the path between src and dst.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
src | yes | String | The value given to the source zone's "id" attribute |
dst | yes | String | The value given to the destination zone's "id" attribute. |
gw_src | yes | String | The value given to the source gateway's "id" attribute; this can be any host or router within the src zone |
gw_dst | yes | String | The value given to the destination gateway's "id" attribute; this can be any host or router within the dst zone |
| symmetrical | no | YES| NO (Default: YES) | If this route is symmetric, the opposite route (from dst to src) will also be declared implicitly. |
<bypasszoneRoute src="my_cluster_1" dst="my_cluster_2" gw_src="my_cluster_1_router" gw_dst="my_cluster_2_router"> <link_ctn id="link_tmp"/> </bypasszoneroute>
This example shows that link link_tmp
(definition not displayed here) directly connects the router my_cluster_1_router
in the source cluster to the router my_cluster_2_router
in the destination router. Additionally, as the symmetrical
attribute was not given, this route is presumed to be symmetrical.
As said before, once you choose a model, it (most likely; the constant network model, for example, doesn't) calculates routes for you. But maybe you want to define some of your routes, which will be specific. You may also want to bypass some routes defined in lower level zone at an upper stage: bypassRoute is the tag you're looking for. It allows to bypass routes defined between host/router. The principle is the same as route: bypassRoute contains list of links references of links that are in the path between src and dst.
Attribute name | Mandatory | Values | Description |
---|---|---|---|
src | yes | String | The value given to the source zone's "id" attribute |
dst | yes | String | The value given to the destination zone's "id" attribute. |
| symmetrical | no | YES | NO (Default: YES) | If this route is symmetric, the opposite route (from dst to src) will also be declared implicitly. |
<bypassRoute src="host_1" dst="host_2"> <link_ctn id="link_tmp"/> </bypassRoute>
This example shows that link link_tmp
(definition not displayed here) directly connects host host_1
to host host_2
. Additionally, as the symmetrical
attribute was not given, this route is presumed to be symmetrical.
Let's say you have an zone named zone_Big that contains two other zone, zone_1 and zone_2. If you want to make a host (h1) from zone_1 with another one (h2) from zone_2 then you'll have to proceed as follows:
As said before, there are mainly 2 tags for routing:
As we are dealing with routes between zone, it means that those we'll have some definition at zone_Big level. Let consider zone_1 contains 1 host, 1 link and one router and zone_2 3 hosts, 4 links and one router. There will be a central router, and a cross-like topology. At the end of the crosses arms, you'll find the 3 hosts and the router that will act as a gateway. We have to define routes inside those two zone. Let say that zone_1 contains full routes, and zone_2 contains some Floyd routing (as we don't want to bother with defining all routes). As we're using some shortest path algorithms to route into zone_2, we'll then have to define some route to gives some topological information to SimGrid. Here is a file doing it all:
<zone id="zone_Big" routing="Dijkstra"> <zone id="zone_1" routing="Full"> <host id="zone_1_host1" speed="1000000000"/> <link id="zone_1_link" bandwidth="1250000000" latency="5E-4"/> <router id="zone_1_gateway"/> <route src="zone_1_host1" dst="zone_1_gateway"> <link_ctn id="zone_1_link"/> </route> </zone> <zone id="zone_2" routing="Floyd"> <host id="zone_2_host1" speed="1000000000"/> <host id="zone_2_host2" speed="1000000000"/> <host id="zone_2_host3" speed="1000000000"/> <link id="zone_2_link1" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link2" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link3" bandwidth="1250000000" latency="5E-4"/> <link id="zone_2_link4" bandwidth="1250000000" latency="5E-4"/> <router id="central_router"/> <router id="zone_2_gateway"/> <!-- routes providing topological information --> <route src="central_router" dst="zone_2_host1"><link_ctn id="zone_2_link1"/></route> <route src="central_router" dst="zone_2_host2"><link_ctn id="zone_2_link2"/></route> <route src="central_router" dst="zone_2_host3"><link_ctn id="zone_2_link3"/></route> <route src="central_router" dst="zone_2_gateway"><link_ctn id="zone_2_link4"/></route> </zone> <link id="backbone" bandwidth="1250000000" latency="5E-4"/> <zoneroute src="zone_1" dst="zone_2" gw_src="zone_1_gateway" gw_dst="zone_2_gateway"> <link_ctn id="backbone"/> </zoneroute> </zone>
The following tags can be used inside a <platform> tag even if they are not directly describing the platform:
Adding configuration flags into the platform file is particularly useful when the described platform is best used with specific flags. For example, you could finely tune SMPI in your platform file directly.
Attribute | Values | Description |
---|---|---|
id | String (optional) | This optional identifier is ignored by SimGrid |
Included tags: <prop> to specify a given configuration item (see options).
Any such configuration must be given at the very top of the platform file.
Example
<?xml version='1.0'?> <!DOCTYPE platform SYSTEM "http://simgrid.gforge.inria.fr/simgrid.dtd"> <platform version="4"> <config> <prop id="maxmin/precision" value="0.000010" /> <prop id="cpu/optim" value="TI" /> <prop id="network/model" value="SMPI" /> <prop id="smpi/bw-factor" value="65472:0.940694;15424:0.697866;9376:0.58729" /> </config> <zone id="zone0" routing="Full"> ...
Defines a user-defined property, identified with a name and having a value. You can specify such properties to most kind of resources: <zone>, <host>, <storage>, <cluster> and <link>. These values can be retrieved at runtime with MSG_zone_property() or simgrid::s4u::NetZone::property(), or similar functions.
Attribute | Values | Description |
---|---|---|
id | String (mandatory) | Identifier of this property. Must be unique for a given property holder, eg host or link. |
value | String (mandatory) | Value of this property; The semantic is completely up to you. |
Included tags: none.
Both tags are an alternate way to pass files containing information on availability, state etc. to an entity. (See also How to model churn). Instead of referring to the file directly in the host, link, or cluster tag, you proceed by defining a trace with an id corresponding to a file, later a host/link/cluster, and finally using trace_connect you say that the file trace must be used by the entity.
<zone id="zone0" routing="Full"> <host id="bob" speed="1000000000"/> </zone> <trace id="myTrace" file="bob.trace" periodicity="1.0"/> <trace_connect trace="myTrace" element="bob" kind="POWER"/>
trace_connect
must come after the elements trace
and host
, as both the host and the trace definition must be known when trace_connect
is parsed; the order of trace
and host
is arbitrary.trace
attributesAttribute name | Mandatory | Values | Description |
---|---|---|---|
id | yes | String | Identifier of this trace; this is the name you pass on to trace_connect . |
file | no | String | Filename of the file that contains the information - the path must follow the style of your OS. You can omit this, but then you must specifiy the values inside of <trace> and </trace> - see the example below. |
trace_periodicity | yes | String | This is the same as for hosts (see there for details) |
Here is an example of trace when no file name is provided:
<trace id="myTrace" periodicity="1.0"> 0.0 1.0 11.0 0.5 20.0 0.8 </trace>
trace_connect
attributes| Attribute name | Mandatory | Values | Description | | ------------— | ------— | -------------------— | --------— | | kind | no | HOST_AVAIL|POWER|
LINK_AVAIL|BANDWIDTH|LATENCY (Default: HOST_AVAIL) | Describes the kind of trace. | | trace | yes | String | Identifier of the referenced trace (specified of the trace's id
attribute) | | element | yes | String | The identifier of the referenced entity as given by its id
attribute |
Now you should know at least the syntax and be able to create a platform by your own. However, after having ourselves wrote some platforms, there are some best practices you should pay attention to in order to produce good platform and some choices you can make in order to have faster simulations. Here's some hints and tips, then.
Most platform files that we ship are in the examples/platforms
folder. The good old grep
tool can find the examples you need when wondering on a specific XML tag. Here is an example session searching for trace_connect:
% cd examples/platforms % grep -R -i -n --include="*.xml" "trace_connect" . ./two_hosts_platform_with_availability_included.xml:26:<trace_connect kind="SPEED" trace="A" element="Cpu A"/> ./two_hosts_platform_with_availability_included.xml:27:<trace_connect kind="HOST_AVAIL" trace="A_failure" element="Cpu A"/> ./two_hosts_platform_with_availability_included.xml:28:<trace_connect kind="SPEED" trace="B" element="Cpu B"/> ./two_hosts.xml:17: <trace_connect trace="Tremblay_power" element="Tremblay" kind="SPEED"/>
This is actually a good idea to search for a better platform file, that better fit the need of your study. To be honest, the provided examples are not representative of anything. They exemplify our XML syntax, but that's all. small_platform.xml for example was generated without much thought beyond that.
The best thing to do when possible is to write your own platform file, that model the platform on which you run your code. For that, you could use our calibration scripts. This leads to very good fits between the platform, the model and the needs. The g5k.xml example resulted of such an effort, which also lead to an ongoing attempt to automatically extract the SimGrid platform from the Grid'5000 experimental platform. But it's hard to come up with generic models. Don't take these files too seriously. Actually, you should always challenge our models and your instanciation if the accuracy really matters to you (see this discussion).
But such advices only hold if you have a real platform and a real application at hand. It's moot for more abstract studies working on ideas and algorithms instead of technical artefacts. Well, in this case, there unfortunately is nothing better than this old and rusty simulacrum. This project is dormant since over 10 years (and you will have to update the generated platforms with bin/simgrid_update_xml
to use them), but that's the best we have for this right now....
The network zone design allows SimGrid to go fast, because computing route is done only for the set of resources defined in the current zone. If you're using only a big zone containing all resource with no zone into it and you're using Full model, then ... you'll loose all interest into it. On the other hand, designing a binary tree of zone with, at the lower level, only one host, then you'll also loose all the good zone hierarchy can give you. Remind you should always be "reasonable" in your platform definition when choosing the hierarchy. A good choice if you try to describe a real life platform is to follow the zone described in reality, since this kind of trade-off works well for real life platforms.
Users that have looked at some of our platforms may have notice a non-intuitive schema ... Something like that:
<zone id="zone_4" routing="Full"> <zone id="exitzone_4" routing="Full"> <router id="router_4"/> </zone> <cluster id="cl_4_1" prefix="c_4_1-" suffix="" radical="1-20" speed="1000000000" bw="125000000" lat="5E-5" bb_bw="2250000000" bb_lat="5E-4"/> <cluster id="cl_4_2" prefix="c_4_2-" suffix="" radical="1-20" speed="1000000000" bw="125000000" lat="5E-5" bb_bw="2250000000" bb_lat="5E-4"/> <link id="4_1" bandwidth="2250000000" latency="5E-5"/> <link id="4_2" bandwidth="2250000000" latency="5E-5"/> <link id="bb_4" bandwidth="2250000000" latency="5E-4"/> <zoneroute src="cl_4_1" dst="cl_4_2" gw_src="c_4_1-cl_4_1_router" gw_dst="c_4_2-cl_4_2_router"> <link_ctn id="4_1"/> <link_ctn id="bb_4"/> <link_ctn id="4_2"/> </zoneroute> <zoneroute src="cl_4_1" dst="exitzone_4" gw_src="c_4_1-cl_4_1_router" gw_dst="router_4"> <link_ctn id="4_1"/> <link_ctn id="bb_4"/> </zoneroute> <zoneroute src="cl_4_2" dst="exitzone_4" gw_src="c_4_2-cl_4_2_router" gw_dst="router_4"> <link_ctn id="4_2"/> <link_ctn id="bb_4"/> </zoneroute> </zone>
In the zone_4, you have an exitzone_4 defined, containing only one router, and routes defined to that zone from all other zone (as cluster is only a shortcut for an zone, see cluster description for details). If there was an upper zone, it would define routes to and from zone_4 with the gateway router_4. It's just because, as we did not allowed (for performances issues) to have routes from an zone to a single host/router, you have to enclose your gateway, when you have zone included in your zone, within an zone to define routes to it.
SimGrid allows you to use some coordinated-based system, like vivaldi, to describe a platform. The main concept is that you have some peers that are located somewhere: this is the function of the coordinates of the <peer> or <host> tag. There's nothing complicated in using it, here is an example:
<?xml version='1.0'?> <!DOCTYPE platform SYSTEM "http://simgrid.gforge.inria.fr/simgrid.dtd"> <platform version="4"> <zone id="zone0" routing="Vivaldi"> <host id="100030591" coordinates="25.5 9.4 1.4" speed="1.5Gf" /> <host id="100036570" coordinates="-12.7 -9.9 2.1" speed="7.3Gf" /> ... <host id="100429957" coordinates="17.5 6.7 18.8" speed="8.3Gf" /> </zone> </platform>
Coordinates are then used to calculate latency (in microseconds) between two hosts by calculating the distance between the two hosts coordinates with the following formula: distance( (x1, y1, z1), (x2, y2, z2) ) = euclidian( (x1,y1), (x2,y2) ) + abs(z1) + abs(z2)
In other words, we take the euclidian distance on the two first dimensions, and then add the absolute values found on the third dimension. This may seem strange, but it was found to allow better approximations of the latency matrices (see the paper describing Vivaldi).
Note that the previous example defines a routing directly between hosts but it could be also used to define a routing between zone. That is for example what is commonly done when using peers (see Section <peer> (Vivaldi netzones only)).
<?xml version='1.0'?> <!DOCTYPE platform SYSTEM "http://simgrid.gforge.inria.fr/simgrid.dtd"> <platform version="4"> <zone id="zone0" routing="Vivaldi"> <peer id="peer-0" coordinates="173.0 96.8 0.1" speed="730Mf" bw_in="13.38MBps" bw_out="1.024MBps" lat="500us"/> <peer id="peer-1" coordinates="247.0 57.3 0.6" speed="730Mf" bw_in="13.38MBps" bw_out="1.024MBps" lat="500us" /> <peer id="peer-2" coordinates="243.4 58.8 1.4" speed="730Mf" bw_in="13.38MBps" bw_out="1.024MBps" lat="500us" /> </zone> </platform>
In such a case though, we connect the zone created by the peer tag with the Vivaldi routing mechanism. This means that to route between zone1 and zone2, it will use the coordinates of router_zone1 and router_zone2. This is currently a convention and we may offer to change this convention in the DTD later if needed. You may have noted that conveniently, a peer named FOO defines an zone named FOO and a router named router_FOO, which is why it works seamlessly with the peer tag.
Choosing wisely the routing model to use can significantly fasten your simulation/save your time when writing the platform/save tremendous disk space. Here is the list of available model and their characteristics (lookup: time to resolve a route):
Each routing model automatically adds a loopback link for each declared host, i.e., a network route from the host to itself, if no such route is declared in the XML file. This default link has a bandwidth of 498 Mb/s, a latency of 15 microseconds, and is not shared among network flows.
If you want to specify the characteristics of the loopback link for a given host, you just have to specify a route from this host to itself with the desired characteristics in the XML file. This will prevent the routing model to add and use the default loopback link.
Actually we did not include switch tag. But when you're trying to simulate a switch, assuming fluid bandwidth models are used (which SimGrid uses by default unless ns-3 or constant network models are activated), the limiting factor is switch backplane bandwidth. So, essentially, at least from the simulation perspective, a switch is similar to a link: some device that is traversed by flows and with some latency and so,e maximum bandwidth. Thus, you can simply simulate a switch as a link. Many links can be connected to this "switch", which is then included in routes just as a normal link.
You have several possibilities, as usual when modeling things. If your cabinets are homogeneous and the intercabinet network negligible for your study, you should just create a larger cluster with all hosts at the same layer.
In the rare case where your hosts are not homogeneous between the cabinets, you can create your cluster completely manually. For that, create an As using the Cluster routing, and then use one <cabinet> for each cabinet. This cabinet tag can only be used an As using the Cluster routing schema, and creating
Be warned that creating a cluster manually from the XML with <cabinet>, <backbone> and friends is rather tedious. The easiest way to retrieve some control of your model without diving into the <cluster> internals is certainly to create one separate <cluster> per cabinet and interconnect them together. This is what we did in the G5K example platform for the Graphen cluster.
It is unfortunately impossible to express the fact that there is more than one routing path between two given hosts. Let's consider the following platform file:
<route src="A" dst="B"> <link_ctn id="1"/> </route> <route src="B" dst="C"> <link_ctn id="2"/> </route> <route src="A" dst="C"> <link_ctn id="3"/> </route>
Although it is perfectly valid, it does not mean that data traveling from A to C can either go directly (using link 3) or through B (using links 1 and 2). It simply means that the routing on the graph is not trivial, and that data do not following the shortest path in number of hops on this graph. Another way to say it is that there is no implicit in these routing descriptions. The system will only use the routes you declare (such as <route src="A" dst="C"><link_ctn id="3"/></route>), without trying to build new routes by aggregating the provided ones.
You are also free to declare platform where the routing is not symmetrical. For example, add the following to the previous file:
<route src="C" dst="A"> <link_ctn id="2"/> <link_ctn id="1"/> </route>
This makes sure that data from C to A go through B where data from A to C go directly. Don't worry about realism of such settings since we've seen ways more weird situation in real settings (in fact, that's the realism of very regular platforms which is questionable, but that's another story).