Differences

This shows you the differences between two versions of the page.

--- grid [2017/10/16 17:57]
popel [Advanced usage]
+++ grid [2018/03/08 18:25]
vidra Add info about profiling RAM by `time -v`
@@ Line 55: / Line 55: @@
   [ -f ~/.bashrc ] && source ~/.bashrc
+Make sure you have correctly configured locale (otherwise ''qrsh'' may not show accented letters in ''less'' and you may get errors when printing utf8 on stdout/stderr from your script in ''qsub''). For example add the following line to your ''~/.bashrc'':
+  export LC_ALL=en_US.UTF-8
 ===== Basic usage =====
@@ Line 119: / Line 122: @@
     * If your job needs more than one CPU (on a single machine) for most of the time, reserve the given number of CPU cores (and SGE slots) with <code>qsub -pe smp <number-of-CPU-cores></code> As you can see in [[#List of Machines]], the maximum is 32 cores. If your job needs e.g. up to 110% CPU most of the time and just occasionally 200%, it is OK to reserve just one core (so you don't waste). TODO: when using ''-pe smp -l mf=8G,amf=8G,h_vmem=12G'', which memory limits are per machine and which are per core?
     * If you are sure your job needs less than 1GB RAM, then you can skip this. Otherwise, if you need e.g. 8 GiB, you must always use ''qsub'' (or ''qrsh'') with ''-l mem_free=8G''. You should specify also ''act_mem_free'' with the same value and ''h_vmem'' with possibly a slightly bigger value. See [[#memory]] for details. TL;DR: <code>qsub -l mem_free=8G,act_mem_free=8G,h_vmem=12G</code>
-  * Be kind to your colleagues. If you are going to submit jobs that effectively occupy **more than one fifth of our cluster for more than several hours**, check if the cluster is free (with ''qstat -g c'' or ''qstat -u \*'') and/or ask your colleagues if they don't plan to use the cluster intensively in the near future. Note that if you allocate one slot (CPU core) on a machine, but (almost) all its RAM, you have effectively occupied the whole machine and all its cores. If you are submitting **more than 100 jobs**, consider using setting them a low priority (e.g. ''-p -1024'', see below) or use [[#qunhold]]. Don't submit more than ca 2000 jobs at once (this can overload the SGE).
+  * Be kind to your colleagues. If you are going to submit jobs that effectively occupy **more than one fifth of our cluster for more than several hours**, check if the cluster is free (with ''qstat -g c'' or ''qstat -u \*'') and/or ask your colleagues if they don't plan to use the cluster intensively in the near future. Note that if you allocate one slot (CPU core) on a machine, but (almost) all its RAM, you have effectively occupied the whole machine and all its cores. If you are submitting **more than 100 jobs**, consider using setting them a low priority (e.g. ''-p -1024'', see below) or use [[#qunhold]].
+  * **Don't submit more than ca 5000 jobs at once**, even if you make sure that at most 100 are running/waiting and the rest is in the //hold// state (e.g. using ''qunhold''). More than 5000 jobs in the queue can overload the SGE, so then no one can execute ''qstat'' (or it takes too long).
@@ Line 129: / Line 133: @@
   * **act_mem_free** (or amf) is a ÚFAL-specific option, which specifies the real amount of free memory (at the time of scheduling). You can specify it when submitting a job and it will be scheduled to a machine with at least this amount of memory free. In an ideal world, where no jobs are exceeding their ''mem_free'' requirements, we would not need this options. However, in real world it is recommended to use this option with the same value as ''mem_free'' to protect your job from failing with out-of-memory error (because of naughty jobs of other users).
   * **h_vmem** is equivalent to setting ''ulimit -v'', i.e. it is a hard limit on the size of virtual memory (see RLIMIT_AS in ''man setrlimit''). If your job exceeds this limit, memory allocation fails (i.e., malloc or mmap will return NULL), and your job will probably crash on SIGSEGV. TODO: according to ''man queue_conf'', the job is killed with SIGKILL, not with SIGSEGV. Note that ''h_vmem'' specifies the maximal size of **allocated_memory, not used_memory**, in other words it is the VIRT column in ''top'', not the RES column. SGE does not use this parameter in any other way. Notably, job scheduling is not affected by this parameter and therefore there is no guarantee that there will be this amount of memory on the chosen machine. The problem is that some programs (e.g. Java with the default setting) allocate much more (virtual) memory than they actually use in the end. If we want to be ultra conservative, we should set ''h_vmem'' to the same value as ''mem_free''. If we want to be only moderately conservative, we should specify something like h_vmem=1.5*mem_free, because some jobs will not use the whole mem_free requested, but still our job will be killed if it allocated much more than declared. The default effectively means that your job has no limits.
-  * It is recommended to **profile your task first**, so you can estimate reasonable memory requirements before submitting many jobs with the same task (varying in parameters which do not affect memory consumption). So for the first time, declare mem_free with much more memory than expected and ssh to a given machine and check ''htop'' (sum all processes of your job) or (if the job is done quickly) check the epilog. When running other jobs of this type, set ''mem_free'' (and ''act_mem_free'' and ''h_vmem'') so you are not wasting resources, but still have some reserve.
+  * It is recommended to **profile your task first** (see [[#profiling]] below), so you can estimate reasonable memory requirements before submitting many jobs with the same task (varying in parameters which do not affect memory consumption). So for the first time, declare mem_free with much more memory than expected and ssh to a given machine and check ''htop'' (sum all processes of your job) or (if the job is done quickly) check the epilog. When running other jobs of this type, set ''mem_free'' (and ''act_mem_free'' and ''h_vmem'') so you are not wasting resources, but still have some reserve.
   * **s_vmem** is similar to ''h_vmem'', but instead of SIGSEGV/SIGKILL, the job is sent a SIGXCPU signal which can be caught by the job and exit gracefully before it is killed. So if you need it, set ''s_vmem'' to a lower value than ''h_vmem'' and implement SIGXCPU handling and cleanup.
@@ Line 138: / Line 142: @@
 This way your job is submitted to the Troja queue. The default is ''ms-all.q''. You can also use e.g.
 ''-q '(troja*|ms*)''' to submit on any machine in those two queues (but **don't use ''-q '*'''** as this includes also [[:gpu|gpu.q]]),
-''-q '*@hector[14]''' to submit on hecor1 or hector4,
+''-q '*@hector[14]''' to submit on hector1 or hector4,
 ''-q '[tm]*@!(hector*|iridium)''' to submit on any troja/ms machine except hectors and iridium.
 However, usually you should specify just the queue (troja-all.q vs. ms-all.q), not a particular machine, and instead use ''-l'' to specify the needed resources in a general way.
@@ Line 145: / Line 149: @@
 See ''man complex'' (run it on lrc or sol machines) for a list of possible resources you may require (in addition to ''mem_free'' etc. discussed above).
-''qsub **-p** -100''
+''qsub **-p** -99''
-Define a priority of your job as a number between -1024 and 0. Only SGE admins may use a number higher than 0. The default is 0, i.e. the highest possible priority. SGE uses the priority to decide when to start which pending job in the queue (it computes a real number called ''prior'', which is reported in ''qstat'', which grows as the job is waiting in the queue). Note that once a job is started, you cannot "unschedule" it, so from that moment on, it is irrelevant what was its priority.
+Define a priority of your job as a number between -1024 and 0. Only SGE admins may use a number higher than 0. In January 2018, we changed the default to -100 (it used to be 0). SGE uses the priority to decide when to start which pending job in the queue (it computes a real number called ''prior'', which is reported in ''qstat'', which grows as the job is waiting in the queue). Note that once a job is started, you cannot "unschedule" it, so from that moment on, it is irrelevant what was its priority. You can ask for a higher priority (-99...0) if your job is urgent and/or will finish soon and you want to skip your colleagues' jobs in the queue. You should ask for lower priority (-1024..-101) if you submit many jobs at once or if the jobs are not urgent.
 ''qsub **-o** LOG.stdout **-e** LOG.stderr''
@@ Line 208: / Line 212: @@
   # immediately on the remote machine
   history -a;
-  # setup the working directory by setting WD
+  # setup the working directory by setting WD, delete possible ".nfs/"
-  ssh -X -Y -C -t $@ "WD='$PWD' /bin/bash --login -i";
+  ssh -X -Y -C -t $@ "WD='${PWD/.nfs\//}' /bin/bash --login -i";
 }
@@ Line 243: / Line 247: @@
   * ''-tc j'' run at most j jobs simultaneously
   * ''-hold_jid_ad comma_separated_job_list'' array jobs that must finish before this job starts; task //i// of the current job depends only on task //i// of the specified jobs
+=== Delete many jobs at once ===
+For deleting all your jobs use ''qdel \*''. For deleting array jobs (see above), you can use comma-separated ranges of task ids, e.g. ''qdel -t 1,10,50-100''.
+For deleting a range of (normal) job ids, you can use bash expansion (as an alternative to ''seq''), e.g. ''qdel {17979..18028}''.
 === Ssh to random sol ===
 Ondřej Bojar suggests to add the following alias to your .bashrc (cf. [[#sshcwd]]):
 <code>alias cluster='comp=$(($RANDOM /4095 +1)); ssh -o "StrictHostKeyChecking no" sol$comp'</code>
 ===== Job monitoring =====
@@ Line 261: / Line 270: @@
   * ''cat /SGE/REPORTER/LRC-UFAL/stats/userlist.weight'' -- all users sorted according to their activity (number of submitted jobs × their average duration), updated each night
   * [[http://ufaladm2/munin/ufal.hide.ms.mff.cuni.cz/lrc-headnode.ufal.hide.ms.mff.cuni.cz/index.html|Munin: graph of cluster usage by day and user]] and  [[http://ufaladm2/munin/ufal.hide.ms.mff.cuni.cz/apophis.ufal.hide.ms.mff.cuni.cz/index.html|Munin monitoring of Apophis disk server]] (both accessible only from ÚFAL network)
+===== Profiling =====
+As stated above, you should always specify the exact memory limits when running your tasks, so that you neither waste RAM nor starve others of memory by using more that you requested. However, memory requirements can be difficult to estimate in advance. That's why you should profile your tasks first.
+A simple method is to run the task and observe the memory usage reported in the epilog, but SGE may not record transient allocations. The kernel does, though, and you can view the exact stats by looking into /proc/$PID/status while the task is running.
+You can still miss allocations made shortly before the program exits – which often happens when trying to debug why your program gets killed by SGE after exhausting the reserved space. To record these, use ''/usr/bin/time -v'' (the actual binary, not the shell-builtin command ''time''). Be aware that unlike the builtin, it cannot measure shell functions and behaves differently on pipelines.
+Obtaining peak usage of multiprocess applications is trickier. Detached and backgrounded processes are ignored completely by ''time -v'' and you get the maximum footprint of any children, not the sum of all maximal footprints nor the largest footprint in any instant.
+If you program in C and need to know the peak memory usage of your children, you can also use the ''wait4()'' syscall and calculate the stats yourself.
 ===== Other =====

[ Back to the navigation ] [ Back to the content ]

Institute of Formal and Applied Linguistics Wiki

Differences