Hi all,

I know that we're leaving the setup the revolves around fyodor soon, but 
I wanted to point out some performance issues that we're currently 
having as they might be able to help us tune the new configuration.  
Specifically, I think that I pointed out some issues before that seemed 
to be caused by slow disk I/O.  Today I felt some delays while using 
fyodor that were confirmed by the vmstat 5 command below:

At about 11:07 EST I got:

procs -----------memory---------- ---swap-- -----io---- -system-- 
----cpu----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy 
id wa
 1  0 866336  41224 150228 234948    0    0     2   518  174  182 11  2 
87  0
 0  0 866336  37544 150364 235084  679    0   704   513  245  313 28  0 
71  0
 2  0 866336  31636 150528 235132    0    0    20   257  165  204 29  3 
68  0
 1 11 866336  36428 150552 235176    0    0     9   572  305  220  8  1 
91  0
 0  0 866336  37072 150652 235196    0    0     2   515  188  245 14  2 
85  0
 0  0 866336  37008 150704 235208    0    0     0   213  130  151 21  1 
78  0

At one point note that there were 11 processes in the "blocked" state, 
which seems to me pretty high for the machine that we have running.  At 
this time, I didn't see spamassassin or a backup process running.  My 
intuition is that the slowdowns are caused by the slow disk subsystem 
using software RAID and normal usage patterns.  Since we also have 
software RAID setup currently on mire, I want to make sure that we're 
not killing our performance before we even move to the new setup.  
Perhaps if we do have to go with software RAID on the front-end system, 
we can mitigate the performance penalty by not using it for AFS caching 
or through some other simple tuning.

At any rate, I would just like to suggest that we do some performance 
tests on the new setup before we migrate most of the current users to 
the new infrastructure.  I added a bullet on performance testing to the 
to-do list because I don't want to overlook this issue, and I think that 
some fairly good performance tests can be done to eliminate the most 
detrimental bottlenecks before we call the new system production-ready.  
>From testing that I've done in the past, I've found that while it's 
really hard to mimic real-world traffic on a test system, there are 
tools that are pre-made, and others that can be written to do a good job 
of finding out what the breaking point of a system will be.

Finally, although I have resigned my formal sysadmin duties, I don't 
want to just toss this task out there for someone else to handle.  If 
you want to let me know when the systems are ready for this stage before 
they're tossed into production, I'll try to spend some doing whatever is 
necessary to help the current admins tune the new configuration.

Best,

Justin

_______________________________________________
HCoop-SysAdmin mailing list
[email protected]
http://hcoop.net/cgi-bin/mailman/listinfo/hcoop-sysadmin

Reply via email to