[
https://issues.apache.org/jira/browse/MESOS-5342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15276194#comment-15276194
]
Chris commented on MESOS-5342:
------------------------------
implementation has been posted to review board. successfully passed 'make
check'. requires use of the hwloc library to perform machine hardware topology
discovery and cpu binding. updates were made to configure.ac and Makefile.am.
implementation is a new "device" under the cgroups isolator directory called
"hwloc". Implementation detects topology, computes total number of cores
required by the container (also checks if the container requires gpu). if the
container requires gpu, the topology information is used to find the "closest"
cores based on latency. If the container only requires cpu, a histogram of task
assignment to cores is checked. If the histogram is "empty" (all cores have a
value of 1.0) then a random core is selected and the latency matrix is used to
find cores that are "closest" to the random core. The histogram is updated. If
the histogram is "not empty" then a greedy submodular subset selection
algorithm is used to select N cores using the latency matrix and a "per-core"
cost value. The "per-core" cost value is a normalized version of the histogram
divided by the number of processing units available on each core. Greedy
submodular subset selection algorithms use a "diminishing returns property" to
find an optimal subset of items under a knapsack constraint.
When the list of cores is returned, a bit vector representing a cpuset is bound
to the container's pid_t. When the container is cleaned up, the histogram is
updated by reducing the current task counts on each core assigned to the pid_t
by -1.0.
> CPU pinning/binding support for CgroupsCpushareIsolatorProcess
> --------------------------------------------------------------
>
> Key: MESOS-5342
> URL: https://issues.apache.org/jira/browse/MESOS-5342
> Project: Mesos
> Issue Type: Improvement
> Components: cgroups, containerization
> Affects Versions: 0.28.1
> Reporter: Chris
>
> The cgroups isolator currently lacks support for binding (also called
> pinning) containers to a set of cores. The GNU/Linux kernel is known to make
> sub-optimal core assignments for processes and threads. Poor assignments
> impact program performance, particularly in the case of applications
> requiring GPU resources.
> Most cluster management systems from the HPC community (SLURM) provide both
> cgroup isolation and cpu binding. This feature would provide similar
> capabilities. The current interest in supporting Intel's Cache Allocation
> Technology will require making choices about where container's are going to
> run on the mesos-agent's processor(s) - this feature is a step toward
> developing a robust solution.
> The improvement in this JIRA ticket will handle hardware topology detection,
> track container-to-core utilization in a histogram, and use a mathematical
> optimization technique to select cores for container assignment based on
> latency and the container-to-core utilization histogram.
> For GPU tasks, the improvement will prioritize selection of cores based on
> latency between the GPU and cores in an effort to minimize copy latency.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
