Grid5000:Gotchas: Difference between revisions

From Grid5000
Jump to navigation Jump to search
 
(22 intermediate revisions by 8 users not shown)
Line 1: Line 1:
{{Maintainer|Lucas Nussbaum}}
{{Maintainer|Lucas Nussbaum}}
{{Portal|User}}
{{Portal|User}}
{{Note|text=For a more upto-date list of Gotchas, see https://www.grid5000.fr/status/artifact/}}
{{Note|text=For a more up to-date list of Gotchas, see https://www.grid5000.fr/status/artifact/}}
This page documents various [http://en.wikipedia.org/wiki/Gotcha_(programming) ''gotchas''] (counter-intuitive features of Grid'5000) that could affect users' experiments in surprising ways.
This page documents various [http://en.wikipedia.org/wiki/Gotcha_(programming) ''gotchas''] (counter-intuitive features of Grid'5000) that could affect users' experiments in surprising ways.


== Network ==
== Network ==
Global and per sites network documentation can be found on [[Grid5000:Network]] page.
=== Topology of ethernet networks ===
=== Topology of ethernet networks ===
Most (large) clusters have a hierarchical ethernet topology, because ethernet switchs with a large number of ports are too expensive. A good example of such a hierarchical topology is the [[Orsay:Network]], where nodes are first connected to 18 24-port switches, which are themselves connected to the central Cisco Catalyst 6509 switch. When doing experiments using the ethernet network intensively, it is a good idea to request nodes on the same switch, using e.g <tt>oarsub -l switch=1/nodes=5</tt>, or to request nodes connected to specific switch using e.g <tt>oarsub -p "switch='cisco2'" -l nodes=5</tt>.
 
Most (large) clusters have a hierarchical ethernet topology, because ethernet switchs with a large number of ports are too expensive. A good example of such a hierarchical topology is the [[Rennes:Network]] for the paravance and parasilo clusters, where nodes are connected to 3 different switches. When doing experiments using the ethernet network intensively, it is a good idea to request nodes on the same switch, using e.g <tt>oarsub -l switch=1/nodes=5</tt>, or to request nodes connected to specific switch using e.g <tt>oarsub -p "switch='cisco2'" -l nodes=5</tt>.


=== Performance of ethernet networks ===
=== Performance of ethernet networks ===
Line 13: Line 17:


=== High-performance networks ===
=== High-performance networks ===
The topology of Infiniband and Myrinet networks is generally less surprising, and many of them are non-blocking (the switch can handle the total bandwidth of all ports simultaneously). However, there are some exceptions :
The topology of Infiniband and Omni-Path networks is generally less surprising, two "fat-tree" topologies can be found on the testbed:
* the Infiniband network in Grenoble is hierarchical (see [[Grenoble:Network]]).
* non-blocking (''1:1'') : the number of up-link ports (from leaf switches to top switches) is equal to the number of down-link ports (nodes to leaf switches). Like that, all the nodes can communicate with each others at full-speed.
* in nancy, graphene-144 is connected to the griffon infiniband switch. This was required in order to free a port on the graphene switch, used to connect the two infiniband switchs together. This can impact the performance of your application if you are using all 144 graphene nodes.
* blocking (''2:1''): the number of up-link ports (from leaf switches to top switches) is half the number of down-link ports (nodes to leaf switches). Like that, nodes from the same leaf switch can communicate to each other at full speed, but not with nodes from others leaf switches.


== Compute nodes ==
== Compute nodes ==
All Grid'5000 clusters are supposed to contain homogeneous (identical) sets of nodes, but there are some exceptions.
All Grid'5000 clusters are supposed to contain homogeneous (identical) sets of nodes, but there are some exceptions.
=== Hard disks ===
Due to their high failure rate, hard disks tend to get replaced frequently, and it is not always possible to keep the same model during the whole life of a cluster.


=== Different disks in the Grenoble edel cluster ===
Global and per sites cluster documentation can be found on [[Hardware]] page.
Disks of the Grenoble edel clusters come in two different size : 128 and 64 GB.
 
* 45 disks of 128GB : 1, 5-6, 8-9, 12, 15-16, 19, 21, 23-25, 27-29, 32, 35, 37, 39-43, 45-46, 48-50, 52, 55-57, 59, 61-63, 65-72
=== Hard disks models ===
* 26 disks of 64GB : 2-4, 7, 10-11, 13-14, 17-18, 20, 22, 26, 30-31, 33-34, 36, 38, 44, 47, 51, 53, 58, 60, 64
Due to their high failure rate, hard disks tend to get replaced frequently, and it is not always possible to keep the same model during the whole life of a cluster. If this is important to you, please check exact disk model using the reference API, as storage is described in detail for each node.
 
=== NVMe disks configuration ===
 
Due to many issues, we had to disable "multipath" support for NVMe disks in most of our environments. This is done by passing the "multipath=off" parameter to the <code>nvme_core</code> module and is done in <code class="command">g5k-postinstall</code>.
 
If you need NVMe multipath support, you can deploy any <code>-min</code> environment (e.g. <code>debian11-x64-min</code>) since they do not contain this workaround.
In such environment, the following limitations apply:
 
* <code class="file">/dev/disk/by-path/</code> entries are not created
* <code class="file">/dev/diskX</code> disk aliases are not created, see below
 
If you need to deploy a <code>-min</code> environment and want working disk aliases, you can use the following workaround on the running system to disable multipath support:
 
<pre>
# rmmod nvme nvme_core
# modprobe nvme_core multipath=off
# modprobe nvme
# sleep 10
</pre>
 
=== Missing disk aliases ===
 
We provide disk aliases such as <code class="file">/dev/diskX</code> on nodes, as documented in hardware pages (e.g. [[Grenoble:Hardware#dahu]]).
 
However, for some configurations, we cannot provide these disk aliases:
 
{| class="wikitable"
|-
! scope="col"| Cluster
! scope="col"| Environment
! scope="col"| Missing disk aliases
! scope="col"| Reason
|-
| [[Grenoble:Hardware#drac|drac]]
| <code class="env">centos7-ppc64-min</code>
| All disks
| Non-standard PCI paths
|-
| [[Grenoble:Hardware#troll|troll]]
| <code class="env">*-x64-min</code>
| NVMe disks
| "NVMe multipath" conflicts with udev
|-
| [[Grenoble:Hardware#yeti|yeti-1]]
| <code class="env">*-x64-min</code>
| <code class="file">/dev/disk4</code>
| "NVMe multipath" conflicts with udev
|-
| [[Grenoble:Hardware#servan|servan]]
| <code class="env">*-x64-min</code>
| NVMe disks
| "NVMe multipath" conflicts with udev
|}


== Software ==
== Software ==
* The production environment (the one users get when not deploying) on all compute nodes is identical, with the exception of additional drivers and software to support GPUs, and Myrinet or Infiniband networks on sites where they are available.
* The standard environment (the one users get when not deploying) on all compute nodes is identical for a given architecture (x86-64, arm64 or ppc64), with the exception of additional drivers and software to support GPUs and High Speed networks on sites where they are available.
* The user frontend are identical on all sites.
* The user frontend are identical on all sites.
* The reference environments ({lenny,squeeze}-x64-{min,base,nfs,xen,big}) are identical on all sites.
* The reference environments (*-$arch-{min,base,nfs,big}) are identical on all sites, for a given architecture.
 
Regarding CPU architectures, some differences can be found in environments:
 
{| class="wikitable"
|-
! scope="col"| Feature
! scope="col"| x86-64
! scope="col"| arm64
! scope="col"| ppc64
! scope="col"| env
|-
! scope="row"| Infiniband
| {{Yes}}
| {{Yes}}
| {{Yes}}
| ''base''
|-
! scope="row"| OmniPath
| {{Yes}}
| {{No}}
| {{No}}
| ''base''
|-
! scope="row"| NFS
| {{Yes}}
| {{Yes}}
| {{Yes}}
| ''nfs''
|-
! scope="row"| Ceph
| {{Yes}}
| {{Yes}}
| {{Yes}}
| ''nfs''
|-
! scope="row"| Cuda
| {{Yes}}
| {{No}}
| {{Yes}}
| ''big''
|-
! scope="row" | BeegFS
| {{Yes}}
| {{No}}
| {{No}}
| ''big''
|-
! scope="row" | OpenMPI
| {{Yes}}
| {{Yes}}
| {{Yes}}
| ''big''
|}

Latest revision as of 16:52, 23 August 2023

Note.png Note

For a more up to-date list of Gotchas, see https://www.grid5000.fr/status/artifact/

This page documents various gotchas (counter-intuitive features of Grid'5000) that could affect users' experiments in surprising ways.

Network

Global and per sites network documentation can be found on Grid5000:Network page.

Topology of ethernet networks

Most (large) clusters have a hierarchical ethernet topology, because ethernet switchs with a large number of ports are too expensive. A good example of such a hierarchical topology is the Rennes:Network for the paravance and parasilo clusters, where nodes are connected to 3 different switches. When doing experiments using the ethernet network intensively, it is a good idea to request nodes on the same switch, using e.g oarsub -l switch=1/nodes=5, or to request nodes connected to specific switch using e.g oarsub -p "switch='cisco2'" -l nodes=5.

Performance of ethernet networks

The backplane bandwidth of ethernet switches doesn't usually allow full-speed communications between all the ports of the switch.

High-performance networks

The topology of Infiniband and Omni-Path networks is generally less surprising, two "fat-tree" topologies can be found on the testbed:

  • non-blocking (1:1) : the number of up-link ports (from leaf switches to top switches) is equal to the number of down-link ports (nodes to leaf switches). Like that, all the nodes can communicate with each others at full-speed.
  • blocking (2:1): the number of up-link ports (from leaf switches to top switches) is half the number of down-link ports (nodes to leaf switches). Like that, nodes from the same leaf switch can communicate to each other at full speed, but not with nodes from others leaf switches.

Compute nodes

All Grid'5000 clusters are supposed to contain homogeneous (identical) sets of nodes, but there are some exceptions.

Global and per sites cluster documentation can be found on Hardware page.

Hard disks models

Due to their high failure rate, hard disks tend to get replaced frequently, and it is not always possible to keep the same model during the whole life of a cluster. If this is important to you, please check exact disk model using the reference API, as storage is described in detail for each node.

NVMe disks configuration

Due to many issues, we had to disable "multipath" support for NVMe disks in most of our environments. This is done by passing the "multipath=off" parameter to the nvme_core module and is done in g5k-postinstall.

If you need NVMe multipath support, you can deploy any -min environment (e.g. debian11-x64-min) since they do not contain this workaround. In such environment, the following limitations apply:

  • /dev/disk/by-path/ entries are not created
  • /dev/diskX disk aliases are not created, see below

If you need to deploy a -min environment and want working disk aliases, you can use the following workaround on the running system to disable multipath support:

# rmmod nvme nvme_core
# modprobe nvme_core multipath=off
# modprobe nvme
# sleep 10

Missing disk aliases

We provide disk aliases such as /dev/diskX on nodes, as documented in hardware pages (e.g. Grenoble:Hardware#dahu).

However, for some configurations, we cannot provide these disk aliases:

Cluster Environment Missing disk aliases Reason
drac centos7-ppc64-min All disks Non-standard PCI paths
troll *-x64-min NVMe disks "NVMe multipath" conflicts with udev
yeti-1 *-x64-min /dev/disk4 "NVMe multipath" conflicts with udev
servan *-x64-min NVMe disks "NVMe multipath" conflicts with udev

Software

  • The standard environment (the one users get when not deploying) on all compute nodes is identical for a given architecture (x86-64, arm64 or ppc64), with the exception of additional drivers and software to support GPUs and High Speed networks on sites where they are available.
  • The user frontend are identical on all sites.
  • The reference environments (*-$arch-{min,base,nfs,big}) are identical on all sites, for a given architecture.

Regarding CPU architectures, some differences can be found in environments:

Feature x86-64 arm64 ppc64 env
Infiniband Check.png Check.png Check.png base
OmniPath Check.png Fail.png Fail.png base
NFS Check.png Check.png Check.png nfs
Ceph Check.png Check.png Check.png nfs
Cuda Check.png Fail.png Check.png big
BeegFS Check.png Fail.png Fail.png big
OpenMPI Check.png Check.png Check.png big