[Clusterusers] sluggishness/errors on new fly launches

Wm. Josiah Erikson wjerikson at hampshire.edu
Mon Jul 24 12:06:37 EDT 2017


I am seeing something similar here on Tom's Novelty Search | Distance:
hamming | Super Anagrams - run 0 job (maybe would happen on any job):

-The node is not running out of RAM and there are no errors

-However, before the job actually starts doing anything, there is a long
(5-10 minute or longer) delay after initial memory allocation:


top - 12:04:09 up 2 days, 19:38,  1 user,  load average: 0.00, 0.11, 0.09
Tasks: 714 total,   1 running, 713 sleeping,   0 stopped,   0 zombie
Cpu(s):  0.1%us,  0.1%sy,  0.0%ni, 99.9%id,  0.0%wa,  0.0%hi,  0.0%si, 
0.0%st
Mem:  66012900k total,  3158436k used, 62854464k free,   157252k buffers
Swap:  2047996k total,        0k used,  2047996k free,   609584k cached

   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+ 
COMMAND         
 63928 thelmuth  20   0 18.7g 330m  11m S  0.0  0.5   0:09.82
java             
 63986 thelmuth  20   0 18.7g 327m  11m S  0.0  0.5   0:10.52
java             
 64232 thelmuth  20   0 18.7g 252m  11m S  0.0  0.4   0:06.68
java             
 64053 thelmuth  20   0 18.7g 185m  11m S  0.0  0.3   0:05.85
java             
 64112 thelmuth  20   0 18.7g 182m  11m S  0.0  0.3   0:05.47
java             
 63838 thelmuth  20   0 18.7g 177m  11m S  0.0  0.3   0:05.90
java             
 64172 thelmuth  20   0 18.7g 173m  11m S  0.0  0.3   0:04.80
java             
 63780 thelmuth  20   0 18.7g 169m  11m S  0.0  0.3   0:05.52
java             

Very little CPU is being used - the network isn't sluggish - it's not
out of RAM (though that's a LOT of VIRT RAM for each job to claim... but
Java just kinda seems to do that)... occasionally you'll see one of the
processes use some CPU for a second... and in a few minutes they will
all start chugging away. Weird and I'm not sure what's up at the moment,
but it doesn't seem to exactly harm anything - just weird.

    -Josiah

   




On 7/19/17 10:55 AM, Wm. Josiah Erikson wrote:
>
> I just saw one of Tom's jobs do the same thing (run rack0 nodes out of
> RAM - not sure if any of the jobs actually crashed yet, but still...),
> so I did two things:
>
> 1. Set the per-blade "sh" limit to 8 (previously unlimited)
>
> 2. Set the number of slots on Rack 0 nodes to 8 (previously 16)
>
>     -Josiah
>
>
>
> On 7/19/17 9:40 AM, Wm. Josiah Erikson wrote:
>>
>> Also, some additional notes:
>>
>> 1. I've set the "bash" limit tag to 8 per node. Tom seems to use sh
>> instead of bash, so this won't affect him, I think.
>>
>> 2. If you want to do 4 per node instead, just add the limit tag
>> "4pernode" to a RemoteCmd in the .alf file, like this:
>>
>>     RemoteCmd "bash -c "env blah blah" -tags "4pernode" -service "tom"
>>
>> 3. There's also a 2pernode tag
>>
>>     -Josiah
>>
>>
>>
>> On 7/19/17 9:30 AM, Wm. Josiah Erikson wrote:
>>>
>>> Yeah, it ran out of RAM - probably tried to take 16 jobs
>>> simultaneously. It's got 64GB of RAM, so should be able to do quite
>>> a few. It's currently running at under 1/3 RAM usage with 3 jobs, so
>>> 8 jobs should be safe. I'll limit "bash" limit tag jobs (what yours
>>> default to since you didn't set one explicitly and that's the
>>> command you're running) to 8 per node. I could lower it to 4 if you
>>> prefer. Let me know.
>>>
>>>     -Josiah
>>>
>>>
>>>
>>> On 7/19/17 8:38 AM, Thomas Helmuth wrote:
>>>> Moving today, so I'll be brief.
>>>>
>>>> This sounds like the memory problems we've had on rack 4. There,
>>>> there are just too many cures for the amount of memory, so if too
>>>> many runs start, none get enough memory. That's why rack 4 isn't on
>>>> the Tom or big go service tags. Probably means we should remove
>>>> rack 0 from those tags for now until we figure it out.
>>>>
>>>> I've done runs on rack 4 by limiting each run to 1GB of memory.
>>>>
>>>> Tom
>>>>
>>>> On Jul 19, 2017 7:46 AM, "Lee Spector" <lspector at hampshire.edu
>>>> <mailto:lspector at hampshire.edu>> wrote:
>>>>
>>>>
>>>>     I launched some runs this morning (a couple of hours ago -- I'm
>>>>     in Europe) that behaved oddly. I don't know if this is related
>>>>     to recent upgrades/reconfiguration, or what, but I thought I'd
>>>>     share them here.
>>>>
>>>>     First, there were weird delays between launches and the
>>>>     generation of the initial output. I'm accustomed to delays of a
>>>>     couple of minutes at this point, which I think we've determined
>>>>     (or just guessed) are due to the necessity of re-downloading
>>>>     dependencies on fly nodes, which doesn't occur when running on
>>>>     other machines. But these took much longer, maybe 30-45 minutes
>>>>     or longer (not sure of the exact times, but they were on that
>>>>     order). And weirder still, there were long times during which
>>>>     many had launched but only one was producing output and
>>>>     proceeding for a long time, with others really kicking in only
>>>>     after the early one(s) finished. I'd understand this if the
>>>>     non-producing ones weren't launched at all, because they were
>>>>     waiting for free slots.... but that's not what was happening.
>>>>
>>>>     And then one crashed, on compute-0-1 (this was launched with
>>>>     the "tom" service tag, using the run-fly script in Clojush),
>>>>     with nothing in the .err file (from standard error) but the
>>>>     following in the .out file (from standard output):
>>>>
>>>>     Producing offspring...
>>>>     Java HotSpot(TM) 64-Bit Server VM warning: INFO:
>>>>     os::commit_memory(0x00000006f0e00000, 1059061760, 0) failed;
>>>>     error='Cannot allocate memory' (errno=12)
>>>>     #
>>>>     # There is insufficient memory for the Java Runtime Environment
>>>>     to continue.
>>>>     # Native memory allocation (malloc) failed to allocate
>>>>     1059061760 bytes for committing reserved memory.
>>>>     # An error report file with more information is saved as:
>>>>     # /home/lspector/runs/july19a-9235/Clojush/hs_err_pid57996.log
>>>>
>>>>      -Lee
>>>>
>>>>     --
>>>>     Lee Spector, Professor of Computer Science
>>>>     Director, Institute for Computational Intelligence
>>>>     Hampshire College, Amherst, Massachusetts, 01002, USA
>>>>     lspector at hampshire.edu <mailto:lspector at hampshire.edu>,
>>>>     http://hampshire.edu/lspector/, 413-559-5352 <tel:413-559-5352>
>>>>
>>>>     _______________________________________________
>>>>     Clusterusers mailing list
>>>>     Clusterusers at lists.hampshire.edu
>>>>     <mailto:Clusterusers at lists.hampshire.edu>
>>>>     https://lists.hampshire.edu/mailman/listinfo/clusterusers
>>>>     <https://lists.hampshire.edu/mailman/listinfo/clusterusers>
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Clusterusers mailing list
>>>> Clusterusers at lists.hampshire.edu
>>>> https://lists.hampshire.edu/mailman/listinfo/clusterusers
>>>
>>> -- 
>>> Wm. Josiah Erikson
>>> Assistant Director of IT, Infrastructure Group
>>> System Administrator, School of CS
>>> Hampshire College
>>> Amherst, MA 01002
>>> (413) 559-6091
>>>
>>>
>>> _______________________________________________
>>> Clusterusers mailing list
>>> Clusterusers at lists.hampshire.edu
>>> https://lists.hampshire.edu/mailman/listinfo/clusterusers
>>
>> -- 
>> Wm. Josiah Erikson
>> Assistant Director of IT, Infrastructure Group
>> System Administrator, School of CS
>> Hampshire College
>> Amherst, MA 01002
>> (413) 559-6091
>
> -- 
> Wm. Josiah Erikson
> Assistant Director of IT, Infrastructure Group
> System Administrator, School of CS
> Hampshire College
> Amherst, MA 01002
> (413) 559-6091

-- 
Wm. Josiah Erikson
Assistant Director of IT, Infrastructure Group
System Administrator, School of CS
Hampshire College
Amherst, MA 01002
(413) 559-6091

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.hampshire.edu/pipermail/clusterusers/attachments/20170724/cefbca00/attachment.html>


More information about the Clusterusers mailing list