On Tue, Jun 22, 2010 at 5:00 PM, tester <solaris.ident...@gmail.com> wrote:
> Hi,
>
> Sorry to bring up an old thread. I need some general guidance on how to use
> dtrace during loads that is greater than normal. Again using on T2+ systemss
> running with a 1 minute load average of 10-15. Running the below script
> shoots up load to 50-60 range and distorts any test.
T2+ implies a multi-processor system with 8 cores x 8 threads. Let's
assume you don't have a T5440. Load average of 15 on a dual
processors T2+ system is less than 12% busy (15 / (2*8*8)).
>
> profile-397
> /pid == 18900/
> {
> �...@[jstack(40,1024)] = count()
> }
>
> tick-5s {
>
> trunc(@,10);
> printa(@);
> trunc(@);
> }
>
> The other symptoms are the system become unresponsive. Any thoughts?
Is it unresponsive only when dtrace is running or normally? With
recent releases of Solaris, I've found systems to be quite responsive
with a load average that is many times higher than the number of CPU's
(as seen by mpstat - 128 for the typical dual processor T5140/T5240).
It seems highly unlikely that the problem is related to being short on
CPU (again, only at about 12% CPU utilization).
If it is unresponsive or sluggish before you start dtrace, I would
guess that one of the following is the case:
- The machine is short on RAM and is paging. Use vmstat to diagnose.
Look at the "b" column (blocked on I/O) and paging related columns
such as sr (scan rate). You would see things as being extremely
sluggish (e.g. when executing a command) because the disk reads needed
to load the commands and related libraries are getting queue behind
the IO requests for paging.
- The network is having troubles. Look for a duplex mismatch or
non-zero values in:
kstat -p e1000g | nawk '$NF != 0 && $0 ~ /(err|drop|fail)/'
- There is some other I/O problem. Does iostat -En show hard errors
on any disk? Does "iostat -xzn 1" show svc_time + wsvc_time over
20ms? How many I/Os are queued and active?
Your question is performance - but you jumped to the conclusion that
dtrace would tell you the answer. It may, but there are likely other
tools that will be helpful with a lot less effort and less system
impact. perf-discuss may be a better list to ask for more help.
--
Mike Gerdts
http://mgerdts.blogspot.com/
_______________________________________________
dtrace-discuss mailing list
dtrace-discuss@opensolaris.org
