Feature-based scheduling is a specific scheduling approach that allows for greater utilization and improved access to resources. It will allow the users to specify exactly what resources they need, allowing the scheduler software flexibility in determining the best resources to run on.
In the past, for example, users without any special resource requests were still required to choose a particular queue, which meant that the job was committed to a particular type of hardware, even when the user did not really care about the specific hardware. The queue structure creates hard boundaries between resource groups, no crossing of boundaries was allowed. However, many times, the users simply didn't care what resources they ran on, as long as they ran. As a result, the feature-based scheduling approach allows the user to specify exactly which resources or features they need, and not specify anything they don't care about.
Similarly, users have only needed a few processors, but a great deal of memory. Since all scheduling was done by processors and time, the best we could recommend was to request more processors, to get the associated memory. This is no longer the case. You can now request the amount of memory separate from the number of processors.
As a general rule, users should request the hardware or features that they need, and not request anything else. The more resources that are requested, the more restricted the scheduler has to be, and the longer it will take to schedule the job. Just for clarity, we'll say this again:
Request the fewest possible resources for your job, in order to minimize the waiting time for the job to start.
What else does this change?
Feature-based scheduling is much more versatile than the old queue based system, which it replaced. We recommend that users learn the feature scheduling approach as soon as possible. If you have any old scripts that used the queue based system (eg. "cuda" or "anynode"), these have been deactivated, and jobs requesting them will not run.
What features are available for scheduling, and how do I use it?
Features that can be used for scheduling fit generically into two major categories: Node-specific features, and Job-specific features, and each of these has it's own syntax. Note that not all combinations of features are available at this time - for example requesting Omni-Path with Infiniband - since the lab does not currently have the specific hardware to fulfill that request.
Resources and features that are node-specific are listed as a part of the resource request. Most users are already familiar with this for the use of the rhel7 or Red Hat 7 request. You may add as many of the following features to your request as you wish.
|Feature name||Description||Syntax example|
|avx||Requests processors that support the avx instruction set.||-C avx|
|avx2||Requests processors that support the avx2 instruction set.||-C avx2|
|avx512||Requests processors that support the avx512 instruction set.||-C avx512|
|fma||Requests processors that support the fma instruction set.||-C fma|
|rhel7||Requests nodes that are running the Red Hat Enterprise 7 Operating System.||-C rhel7|
|skylake||Requests nodes that have Skylake processors.||-C skylake|
|knl||Requests nodes that have Knights Landing (Xeon Phi) processors.||-C knl|
|ib||Requests that the nodes have access to an Infiniband interconnect||-C ib|
|kepler||Requests nodes that have Kepler GPU's.||-C kepler|
|opa||Requests that the nodes have access to an Omni-Path interconnect||-C opa|
|pascal||Requests nodes that have Pascal GPU's.||-C pascal|
Example Usage of Node-specific features
There are several options available to request multiple features for a particular request. For example an ampersand (&) can be used to denote multiple features that are required to be used together. A bar symbol (|) can be used to denote features where either can be used. Note that because some symbols (eg. ! and |) have special meaning in the shell environment, they must be quoted as shown here to have the desired effect.
|Requests that the nodes have Skylake processors with infiniband||-C 'skylake&ib'|
|Requests nodes that have Kepler, or Pascal GPU's (doesn't matter which).||-C 'kepler|pascal'|
|Requests nodes that have Knights Landing processors, and do NOT have infiniband (ib)||-C 'knl&!ib'|
Job-specific features apply to the job as a whole, and are listed in the resource request:
|Feature name||Description||Syntax example||Note|
|nodes||Requests a specific number of nodes||--nodes=4 --ntasks=4 --mem-per-cpu=2G --time=1:00:00||This translates into a total of 4 nodes; if not specified, the system will assume 1 node.|
|ntasks||Requests a total number of processors||--nodes=4 --ntasks=4 --mem-per-cpu=2G --time=1:00:00||This translates into a total of 4 processors across nodes (1 processor per node in this example).|
|time||Requests a total amount of running time for the job||--nodes=1 --ntasks=4 --mem-per-cpu=2G --time=1:00:00||Format is time=hh:mm:ss; if not specified, the system will assume 1 hour|
|mem||Requests a total amount of memory requested by all processes in the job||--nodes=1 --ntasks=4 --mem=16G||This translates into an even division of memory per process (4gb per process in the example)|
|mem-per-cpu||Requests a specific amount of memory per process||--nodes=1 --ntasks=4 --mem-per-cpu=4G||This is equivalent to the "--mem=16G" listed above|
|qos||Requests a specific QOS, or quality of service. This is used primarily for preemption.||--nodes=1 --ntasks=4 --mem-per-cpu=2G --qos=preemptee||The example specifies that this job is preemptable|
The following examples combine several of the above requirements, and include explanation of the specifics.
|--nodes=1 --ntasks=4 --mem-per-cpu=2G -C 'ib&intel' --time=4:30:00||Requesting 4 processors on 1 node, with Infiniband, and Intel processors, for 4.5 hours|
|--nodes=16 --ntasks=8 --mem-per-cpu=2gb -C 'skylake:ib&gpu' time=168:00:00 qos=preemptee||Requesting 16 nodes with 8 Intel Skylake processors each, with Infiniband, with GPGPU's attached, for 168 hours (7 days), and the preemptee QOS|
What resource requests are required?
The purpose of the feature scheduling approach is to allow users to request only the resources or features they care about, and not specify the features they don't care about. However, we recommend as a minimum the use of nodes, ppn, and walltime as shown below. If these are not specified, the system has some defaults it will apply, which may not be what you intended. At the time of this writing, the defaults are 1 node, 1 processor per node, and 1 hour of walltime.
--nodes=10 --ntasks=8 --mem-per-cpu=2gb --time=24:00:00