Jeff/Joseph,
Sorry for taking so long to follow up with this but I believe I've
fixed this issue for good and you should now be able to launch 50+
node clusters without issue. My original feeling was that the SGE
install script was at fault, however, after several hours of digging I
discovered that ssh-keyscan was failing when there were a large number
of nodes. Long story short this meant that passwordless-ssh wasn't
being setup fully for all nodes and so the SGE installer script could
not connect to those nodes to add them to the queue. I found a much
better way to populate the known_hosts file with all the nodes using
paramiko instead of ssh-keyscan which is much faster in this case.
If you haven't already please re-run 'python setup.py install' after
pulling the latest code to test out the latest changes. I've also
updated StarCluster perform the setup on all nodes concurrently using
a thread pool so you should notice it's much faster for larger
clusters. Please let me know if you have issues.
Thanks,
~Justin
On Wed, Mar 16, 2011 at 1:37 PM, Kyeong Soo (Joseph) Kim
<kyeongsoo.kim_at_gmail.com> wrote:
> Justin,
> Please, find attached the said file.
>
> Regards,
> Joseph
>
>
> On Wed, Mar 16, 2011 at 4:38 PM, Justin Riley <jtriley_at_mit.edu> wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> Joseph,
>>
>> Great thanks, can you also send me the /opt/sge6/ec2_sge.conf file please?
>>
>> ~Justin
>>
>> On 03/16/2011 12:29 PM, Kyeong Soo (Joseph) Kim wrote:
>>> Hi Justin,
>>>
>>> Please, find attached the gzipped tar file of the logfiles under
>>> install_logs directory.
>>>
>>> Note that the configuration is for 25-node (1 master and 24 slaves) cluster.
>>>
>>> Below is the time-sorted listing of log files under the same directory:
>>>
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 13:23
>>> execd_install_node024_2011-03-16_13:23:11.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node023_2011-03-16_11:13:37.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node022_2011-03-16_11:13:36.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node021_2011-03-16_11:13:36.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node020_2011-03-16_11:13:32.log
>>> -rw-r--r-- 1 kks kks 18K 2011-03-16 11:13
>>> execd_install_master_2011-03-16_11:13:10.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node017_2011-03-16_11:13:27.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node018_2011-03-16_11:13:27.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node019_2011-03-16_11:13:28.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node016_2011-03-16_11:13:26.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node014_2011-03-16_11:13:25.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node015_2011-03-16_11:13:26.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node012_2011-03-16_11:13:24.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node013_2011-03-16_11:13:25.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node010_2011-03-16_11:13:23.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node011_2011-03-16_11:13:24.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node008_2011-03-16_11:13:22.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node009_2011-03-16_11:13:22.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node006_2011-03-16_11:13:21.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node007_2011-03-16_11:13:21.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node004_2011-03-16_11:13:20.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node005_2011-03-16_11:13:20.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node003_2011-03-16_11:13:19.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node001_2011-03-16_11:13:18.log
>>> -rw-r--r-- 1 kks kks 2.9K 2011-03-16 11:13
>>> execd_install_node002_2011-03-16_11:13:19.log
>>> -rw-r--r-- 1 kks kks 3.1K 2011-03-16 11:13
>>> execd_install_master_2011-03-16_11:13:17.log
>>> -rw-r--r-- 1 kks kks 8.4K 2011-03-16 11:13
>>> qmaster_install_master_2011-03-16_11:13:05.log
>>>
>>> As you can see, the installation of master has been duplicated and it
>>> ended up with master, node001~node023; the top-most log for node024
>>> was for the manual addition through "addnode" command later (i.e., 1
>>> hour 10 mins after).
>>>
>>> Even with this slimmed down version of configurations (compared to the
>>> original 125-node one), the chances that all nodes are properly
>>> installed (i.e., 25 out of 25) were about 50% (last night and this
>>> morning, I tried it about 10 times to set total five of 25-node
>>> clusters).
>>>
>>> Regards,
>>> Joseph
>>>
>>>
>>> On Wed, Mar 16, 2011 at 3:57 PM, Justin Riley <jtriley_at_mit.edu> wrote:
>>> Hi Jeff/Joseph,
>>>
>>> I just requested to up my EC2 instance limit so that I can test things
>>> out at this scale and see what the issue is. In the mean time would you
>>> mind sending me any logs found in /opt/sge6/default/common/install_logs
>>> and also the /opt/sge6/ec2_sge.conf for a failed run?
>>>
>>> Also if this happens again you could try reinstalling SGE manually
>>> assuming all the nodes are up:
>>>
>>> $ starcluster sshmaster mycluster
>>> $ cd /opt/sge6
>>> $ ./inst_sge -m -x -auto ./ec2_sge.conf
>>>
>>> ~Justin
>>>
>>> On 03/15/2011 06:30 PM, Kyeong Soo (Joseph) Kim wrote:
>>>>>> Hi Jeff,
>>>>>>
>>>>>> I experienced the same thing with my 50-node configuration (c1.xlarge).
>>>>>> Out of 50 nodes, only 29 nodes are successfully identified by the SGE.
>>>>>>
>>>>>> Regards,
>>>>>> Joseph
>>>>>>
>>>>>> On Sat, Mar 5, 2011 at 10:15 PM, Jeff White <jeff_at_decide.com> wrote:
>>>>>>> I can frequently reproduce an issue where 'starcluster start' completes
>>>>>>> without error, but not all nodes are added to the SGE pool, which I verify
>>>>>>> by running 'qconf -sel' on the master. The latest example I have is creating
>>>>>>> a 25-node cluster, where only the first 12 nodes are successfully installed.
>>>>>>> The remaining instances are running and I can ssh to them but they aren't
>>>>>>> running sge_execd. There are only install log files for the first 12 nodes
>>>>>>> in /opt/sge6/default/common/install_logs. I have not found any clues in the
>>>>>>> starcluster debug log or the logs inside master:/opt/sge6/.
>>>>>>>
>>>>>>> I am running starcluster development snapshot 8ef48a3 downloaded on
>>>>>>> 2011-02-15, with the following relevant settings:
>>>>>>>
>>>>>>> NODE_IMAGE_ID=ami-8cf913e5
>>>>>>> NODE_INSTANCE_TYPE = m1.small
>>>>>>>
>>>>>>> I have seen this behavior with the latest 32-bit and 64-bit starcluster
>>>>>>> AMIs. Our workaround is to start a small cluster and progressively add nodes
>>>>>>> one at a time, which is time-consuming.
>>>>>>>
>>>>>>> Has anyone else noticed this and have a better workaround or an idea for a
>>>>>>> fix?
>>>>>>>
>>>>>>> jeff
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> StarCluster mailing list
>>>>>>> StarCluster_at_mit.edu
>>>>>>> http://mailman.mit.edu/mailman/listinfo/starcluster
>>>>>>>
>>>>>>>
>>>>>> _______________________________________________
>>>>>> StarCluster mailing list
>>>>>> StarCluster_at_mit.edu
>>>>>> http://mailman.mit.edu/mailman/listinfo/starcluster
>>>
>>>>
>>
>> -----BEGIN PGP SIGNATURE-----
>> Version: GnuPG v2.0.17 (GNU/Linux)
>> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>>
>> iEYEARECAAYFAk2A550ACgkQ4llAkMfDcrlR2gCeOoYMzl9U+z1owIq98JHBgLHi
>> IngAniUwV6nq/hN6/TfxCBu1d2/MO5Ru
>> =tXep
>> -----END PGP SIGNATURE-----
>>
>
> _______________________________________________
> StarCluster mailing list
> StarCluster_at_mit.edu
> http://mailman.mit.edu/mailman/listinfo/starcluster
>
>
Received on Wed Apr 06 2011 - 11:09:24 EDT