CoolThreads and load average -- SUMMARY

Thanks to everyone who answered.  The consensus was that the 5120 has 64 virtual processors, so a load average of 30 is no big deal.  I was used to the rule of thumb that you didn't want to see a load average above 4-5 on the old systems. 

Here are the comments:
--------------------------------------------------------------------

Not sure if you have seen this, but it may provide some insight:

http://www.solarisinternals.com/wiki/index.php/CMT_Utilization

Load average is a figure that approximates the average number of
processes that are either running or waiting to run on a CPU.  Since
this machine has many virtual CPUs, having a load average on the order
of that many CPUs sounds reasonable. 

Load average is not a concern until it exceeds 4 times
the number of CPU threads available to it.

Load Average is the number of jobs in the run queue plus the number of 
jobs on CPU averaged over the time period of 1, 5 and 15 minutes (those 
are the 3 numbers you get when you enter `uptime`). So, if you have a 2 
CPU system with a load average of 30, you are having trouble. If you 
have a system with an apparent 64 CPU's (8 core, 8 threads per core), 
then 30 is no big deal. I haven't managed to get my T5220's cranked up 
yet (still in development), so every time I check uptime, I get a load 
average of 0.00

Back in the day, anything load average over about 4-5 meant

bad performance.  However, with current hardware, this isn't

true any more.  I've seen load averages as high as 98 on some

systems while running backups, and there wasn't any difference

in performance, other than sendmail was all bitter and wouldn't

talk because the load average was above 12.



However, this is very dependent on what the load is - 98 cpu

and memory hogs will likely hurt performance, but a pile of

things stuck in I/O wait isn't a big deal if you have the disk

bandwidth to handle the traffic, so that the backup tasks complete

in a sane amount of time.

You seem to be equating a high load average with bad performance.  As
a rule of thumb, load average is not a concern until it exceeds 4 times
the number of CPU threads available to it.

Within reason, a high load average just means that your machine is earning
its keep by being busy!



-----Original Message-----
From: sunmanagers-bounces@sunmanagers.org [mailto:sunmanagers-bounces@sunmanagers.org] On Behalf Of Dan Smith
Sent: Wednesday, November 19, 2008 11:24
To: sunmanagers@sunmanagers.org
Subject: CoolThreads and load average

Hello Sun Managers:

We are running T5120 boxes for our test, DR, and production databases.
They have been performing very well in supporting Oracle database.  We
are looking at running RMAN backups with compression to save disk space,
tape, and bandwidth.  We run the backup with 32 channels configured in
the RMAN backup.  This throttles the process down to using 50% of the
CPU. This is good. However, we see a load average of 30 during the same
time period.

We will be doing some more testing to determine the effect of the load
average of 30 on the database performance.  In the meantime, have you
seen any comparable phenomena with the CoolThreads servers?  I am
wondering if we can interpret the load average of 30, differently on
these boxes.

Thanks,
Dan Smith
Western Michigan University OIT
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers