From Beocat
Jump to: navigation, search
No edit summary
(Added MPI instructions.)
Line 432: Line 432:
=== CUDA ===
=== CUDA ===
[[CUDA]] is the resource required for GPU computing. We have a very small number of nodes which have GPUs installed. To request one of these nodes, add <tt>-l cuda=true</tt> to your qsub command-line.
[[CUDA]] is the resource required for GPU computing. We have a very small number of nodes which have GPUs installed. To request one of these nodes, add <tt>-l cuda=true</tt> to your qsub command-line.
=== Exclusive ===
Some programs just don't play nicely with others. They will attempt to use all available memory or will try to use all the cores it can use. The way to be a nice neighbor if your program has this problem is to request exclusive use of a node with <tt>-l excl=true</tt>. This can also be useful for benchmarking, where you can be sure that no other jobs are interfering with yours.
== Parallel Jobs ==
There are two ways jobs can run in parallel, ''intra''node and ''inter''node. '''Note: Beocat will not automatically make a job run in parallel.''' Have I said that enough? It's a common misperception.
=== Intranode jobs ===
Intranode jobs are easier to code and can take advantage of many common libraries, such as [http://openmp.org/wp/ OpenMP], or Java's threads. Many times, your program will need to know how many cores you want it to use. Many will use all available cores if not told explicitly otherwise. This can be a problem when you are sharing resources, as Beocat does. To request multiple cores, use the qsub directive '<tt>-pe single ''n''</tt>', where ''n'' is the number of cores you wish to use. If your command can take an environment variable, you can use $nslots to tell how many cores you've been allocated.
=== Internode (MPI) jobs ===
"Talking" between nodes is trickier that talking between cores on the same node. The specification for doing so is called "[[wikipedia:Message_Passing_Interface|Message Passing Interface]]", or MPI. We have [http://www.open-mpi.org/ OpenMPI] installed on Beocat for this purpose. Most programs written to take advantage of large multi-node systems will use MPI. You can tell if you have an MPI-enabled program because its directions will tell you to run '<tt>mpirun ''program''</tt>'. Requesting MPI resources is only mildly more difficult than requesting single-node jobs. Instead of using '<tt>-pe single ''n''</tt>' for your qsub request, you will use one of the following:
{|
|mpi-fill
|This environment will use as many slots on each node as it can until it reaches the number of cores you have requested.
|-
|mpi-spread
|This environment will spread itself out over as many nodes as possible until it reaches the number of cores you have requested.
|-
|mpi-1
|This environment will allocate the slots you've requested 1 per node.
|-
|mpi-2
|This environment will allocate the slots you've requested 2 per node. You must request cores as a multiple of 2
|-
|mpi-4
|This environment will allocate the slots you've requested 4 per node. You must request cores as a multiple of 4
|-
|mpi-8
|This environment will allocate the slots you've requested 8 per node. You must request cores as a multiple of 8
|-
|mpi-10
|This environment will allocate the slots you've requested 10 per node. You must request cores as a multiple of 10
|-
|mpi-12
|This environment will allocate the slots you've requested 12 per node. You must request cores as a multiple of 12
|-
|mpi-16
|This environment will allocate the slots you've requested 16 per node. You must request cores as a multiple of 16
|-
|mpi-80
|This environment will allocate the slots you've requested 80 per node. You must request cores as a multiple of 80
|}
Some quick examples:
<tt>-pe mpi-4 16</tt> will give you 4 chunks of 4 cores apiece. They might all happen to be allocated on the same node (16 cores), on 4 different nodes (4 cores each), on 3 nodes (8 cores on one and 4 cores on the other two), or on 2 nodes (8 cores each).
<tt>-pe mpi-fill 40</tt> will give you 40 cores, but will attempt to get them all on the same node.
<tt>-pe mpi-fill 100</tt> will give you 100 cores, and place them on as few nodes as possible. In this case it's likely you would get a full mage (80 cores) and either part of another mage (the remaining 20 cores) or one of the 20-core elves.
<tt>-pe mpi-spread 40</tt> will give you 40 cores, and will attempt to place each on a separate node.
== Requesting memory for multi-core jobs ==
All memory requests are '''per core'''. One of the more common scenarios is where somebody will need, say 20 cores and 400 GB of memory. So they will make a request like '<tt>-pe single 20, -l mem=400G</tt>' This will never run, because what you are really requesting is 20 cores and 8000GB of memory (20 * 400). Since we have no nodes with 8000 terabytes of memory, the job will never run. In this case, you will divide the 400GB total memory request by the number of cores (20), so the correct command would be '<tt>-pe single 20, -l mem=20G</tt>'.
== Running jobs interactively ==
Some jobs just don't behave like we think they should, or need to be run with somebody sitting at the keyboard and typing in response to the output the computers are generating. Beocat has a facility for this, called 'qrsh'. qrsh uses the exact same command-line arguments as qsub. If no node is available with your resource requirements, qrsh will tell you
Your "qrsh" request could not be scheduled, try again later.
Note that, like qsub, your interactive job will timeout after your allotted time has passed.
== Job Accounting ==
== Job Accounting ==
Some people may find it useful to know what their job did during its run. The qacct tool will read SGE's accounting file and give you summarized or detailed views on jobs that have run within Beocat.
Some people may find it useful to know what their job did during its run. The qacct tool will read SGE's accounting file and give you summarized or detailed views on jobs that have run within Beocat.

Revision as of 16:42, 7 July 2014

Resource Requests

Aside from the time, RAM, and CPU requirements listed on the SGEBasics page, we have several other requestable resources. Generally, if you don't know if you need a particular resource, you should use the default. These can be generated with the command

qconf -sc | awk '{ if ($5 != "NO") { print }}'
name shortcut type relop requestable consumable default urgency
arch a RESTRING == YES NO NONE 0
avx avx BOOL == YES NO FALSE 0
calendar c RESTRING == YES NO NONE 0
cpu cpu DOUBLE >= YES NO 0 0
cpu_flags c_f STRING == YES NO NONE 0
cuda cuda INT <= YES JOB 0 0
display_win_gui dwg BOOL == YES NO 0 0
exclusive excl BOOL EXCL YES YES 0 1000
h_core h_core MEMORY <= YES NO 0 0
h_cpu h_cpu TIME <= YES NO 0:0:0 0
h_data h_data MEMORY <= YES NO 0 0
h_fsize h_fsize MEMORY <= YES NO 0 0
h_rss h_rss MEMORY <= YES NO 0 0
h_rt h_rt TIME <= FORCED NO 0:0:0 0
h_stack h_stack MEMORY <= YES NO 0 0
h_vmem h_vmem MEMORY <= YES NO 0 0
hostname h HOST == YES NO NONE 0
infiniband ib BOOL == YES NO FALSE 0
m_core core INT <= YES NO 0 0
m_socket socket INT <= YES NO 0 0
m_thread thread INT <= YES NO 0 0
m_topology topo RESTRING == YES NO NONE 0
m_topology_inuse utopo RESTRING == YES NO NONE 0
mem_free mf MEMORY <= YES NO 0 0
mem_total mt MEMORY <= YES NO 0 0
mem_used mu MEMORY >= YES NO 0 0
memory mem MEMORY <= FORCED YES 0 0
num_proc p INT == YES NO 0 0
qname q RESTRING == YES NO NONE 0
s_core s_core MEMORY <= YES NO 0 0
s_cpu s_cpu TIME <= YES NO 0:0:0 0
s_data s_data MEMORY <= YES NO 0 0
s_fsize s_fsize MEMORY <= YES NO 0 0
s_rss s_rss MEMORY <= YES NO 0 0
s_rt s_rt TIME <= YES NO 0:0:0 0
s_stack s_stack MEMORY <= YES NO 0 0
s_vmem s_vmem MEMORY <= YES NO 0 0
slots s INT <= YES YES 1 1000
swap_free sf MEMORY <= YES NO 0 0
swap_rate sr MEMORY >= YES NO 0 0
swap_rsvd srsv MEMORY >= YES NO 0 0
swap_total st MEMORY <= YES NO 0 0
swap_used su MEMORY >= YES NO 0 0
virtual_free vf MEMORY <= YES NO 0 0
virtual_total vt MEMORY <= YES NO 0 0
virtual_used vu MEMORY >= YES NO 0 0

The good news is that most of these nobody ever uses. There are a couple of exceptions, though:

Infiniband

First of all, let me state that just because it sounds "cool" doesn't mean you need it or even want it. Infiniband does absolutely no good if running in a 'single' parallel environment. Infiniband is a high-speed host-to-host communication fabric. It is used in conjunction with MPI jobs (discussed below). Several times we have had jobs which could run just fine, except that the submitter requested Infiniband, and all the nodes with Infiniband were currently busy. In fact, some of our fastest nodes do not have Infiniband, so by requesting it when you don't need it, you are actually slowing down your job. To request Infiniband, add -l ib=true to your qsub command-line.

CUDA

CUDA is the resource required for GPU computing. We have a very small number of nodes which have GPUs installed. To request one of these nodes, add -l cuda=true to your qsub command-line.

Exclusive

Some programs just don't play nicely with others. They will attempt to use all available memory or will try to use all the cores it can use. The way to be a nice neighbor if your program has this problem is to request exclusive use of a node with -l excl=true. This can also be useful for benchmarking, where you can be sure that no other jobs are interfering with yours.

Parallel Jobs

There are two ways jobs can run in parallel, intranode and internode. Note: Beocat will not automatically make a job run in parallel. Have I said that enough? It's a common misperception.

Intranode jobs

Intranode jobs are easier to code and can take advantage of many common libraries, such as OpenMP, or Java's threads. Many times, your program will need to know how many cores you want it to use. Many will use all available cores if not told explicitly otherwise. This can be a problem when you are sharing resources, as Beocat does. To request multiple cores, use the qsub directive '-pe single n', where n is the number of cores you wish to use. If your command can take an environment variable, you can use $nslots to tell how many cores you've been allocated.

Internode (MPI) jobs

"Talking" between nodes is trickier that talking between cores on the same node. The specification for doing so is called "Message Passing Interface", or MPI. We have OpenMPI installed on Beocat for this purpose. Most programs written to take advantage of large multi-node systems will use MPI. You can tell if you have an MPI-enabled program because its directions will tell you to run 'mpirun program'. Requesting MPI resources is only mildly more difficult than requesting single-node jobs. Instead of using '-pe single n' for your qsub request, you will use one of the following:

mpi-fill This environment will use as many slots on each node as it can until it reaches the number of cores you have requested.
mpi-spread This environment will spread itself out over as many nodes as possible until it reaches the number of cores you have requested.
mpi-1 This environment will allocate the slots you've requested 1 per node.
mpi-2 This environment will allocate the slots you've requested 2 per node. You must request cores as a multiple of 2
mpi-4 This environment will allocate the slots you've requested 4 per node. You must request cores as a multiple of 4
mpi-8 This environment will allocate the slots you've requested 8 per node. You must request cores as a multiple of 8
mpi-10 This environment will allocate the slots you've requested 10 per node. You must request cores as a multiple of 10
mpi-12 This environment will allocate the slots you've requested 12 per node. You must request cores as a multiple of 12
mpi-16 This environment will allocate the slots you've requested 16 per node. You must request cores as a multiple of 16
mpi-80 This environment will allocate the slots you've requested 80 per node. You must request cores as a multiple of 80

Some quick examples:

-pe mpi-4 16 will give you 4 chunks of 4 cores apiece. They might all happen to be allocated on the same node (16 cores), on 4 different nodes (4 cores each), on 3 nodes (8 cores on one and 4 cores on the other two), or on 2 nodes (8 cores each).

-pe mpi-fill 40 will give you 40 cores, but will attempt to get them all on the same node.

-pe mpi-fill 100 will give you 100 cores, and place them on as few nodes as possible. In this case it's likely you would get a full mage (80 cores) and either part of another mage (the remaining 20 cores) or one of the 20-core elves.

-pe mpi-spread 40 will give you 40 cores, and will attempt to place each on a separate node.

Requesting memory for multi-core jobs

All memory requests are per core. One of the more common scenarios is where somebody will need, say 20 cores and 400 GB of memory. So they will make a request like '-pe single 20, -l mem=400G' This will never run, because what you are really requesting is 20 cores and 8000GB of memory (20 * 400). Since we have no nodes with 8000 terabytes of memory, the job will never run. In this case, you will divide the 400GB total memory request by the number of cores (20), so the correct command would be '-pe single 20, -l mem=20G'.

Running jobs interactively

Some jobs just don't behave like we think they should, or need to be run with somebody sitting at the keyboard and typing in response to the output the computers are generating. Beocat has a facility for this, called 'qrsh'. qrsh uses the exact same command-line arguments as qsub. If no node is available with your resource requirements, qrsh will tell you

Your "qrsh" request could not be scheduled, try again later.

Note that, like qsub, your interactive job will timeout after your allotted time has passed.

Job Accounting

Some people may find it useful to know what their job did during its run. The qacct tool will read SGE's accounting file and give you summarized or detailed views on jobs that have run within Beocat.

qacct

This data can usually be used to diagnose two very common job failures.

Job debugging

It is simplest if you know the job number of the job you are trying to get information on.

# if you know the jobid, put it here:
qacct -j 1122334455
# if you don't know the job id, you can look at your jobs over some number of days in this case the past 14 days:
qacct -o $USER -d 14 -j
My job didn't do anything when it ran!
qname        batch.q             
hostname     mage07.beocat       
group        some_user_users        
owner        some_user              
project      BEODEFAULT          
department   defaultdepartment   
jobname      my_job_script.sh  
jobnumber    1122334455          
...
snipped to save space
...
exit_status  1                   
ru_wallclock 1s
ru_utime     0.030s
ru_stime     0.030s
...
snipped to save space
...
arid         undefined
category     -u some_user -q batch.q,long.q -l h_rt=604800,mem_free=1024.0M,memory=2G

If you look at the line showing ru_wallclock. You can see that it shows 1s. This means that the job started and then promptly ended. This points to something being wrong with your submission script. Perhaps there is a typo somewhere in it.

My job ran but didn't finish!
qname        batch.q             
hostname     scout59.beocat      
group        some_user_users     
owner        some_user           
project      BEODEFAULT          
department   defaultdepartment   
jobname      my_job_script.sh           
jobnumber    1122334455            
...
snipped to save space
...            
slots        1                   
failed       37  : qmaster enforced h_rt, h_cpu, or h_vmem limit
exit_status  0                   
ru_wallclock 21600s
ru_utime     0.130s
ru_stime     0.020s
...
snipped to save space
...
arid         undefined
category     -u some_user -q batch.q,long.q -l h_rt=21600,mem_free=512.0M,memory=1G

If you look at the lines showing failed, ru_wallclock and category we can see some pointers to the issue. It didn't finish because the scheduler (qmaster) enforced some limit. If you look at the category line, the only limit requested was h_rt. So it was a runtime (wallclock) limit. Comparing ru_wallclock and the h_rt request, we can see that it ran until the h_rt time was hit, and then the scheduler enforce the limit and killed the job. You will need to resubmit the job and ask for more time next time.