System Upgrade but poorer performance...?

[steve_586]

Hi,

Have a D330 that has had a D370 Proc & motherboard upgrade and also has had a 200% RAM increase (to 768MB) in order to improve performance. The problem is the sar outputs report the system running less IDLE than it was prior to the upgrade, the server does in fact 'feel' a little slower. I have attatched current kern settings and other info just in case nayone out there has any suggestions. Swap has been increased to 1GB (is that enough?)If any ouputs are required the please let me know. Any ideas would be appreciated on this. Thanks

[Steven E. Protter]

Swap is at your level recommended at twice physical memory. Some say as little as 1.5 times physical memory, so you are a bit short there.

If you have a disk small enough, you might dedicate it to swap, and not have say an Oracle database on the same disk if dedication is not an option.

I'm attaching a more comprehensive background sar script to help you in diagnosis.

SEP

[Patrick Wallek]

It appears to me that your problem is not with CPU power or RAM and you are not swapping so your swap config is fine. The problem appears to be the %WIO. You are doing lots of waiting on IO.

If you are just using the disks internal to the D, then this is not altogether unexpected. I support lots of D boxes and almost all of them exhibit this behavior.

You may be able to do something if you were to attach some type of external storage and try some striping of the disks or something.

[steve_586]

I did look at disk i/o but it didnt really jump out at me as being really high. I have attatched two examples of disk i/o when the server is running at 'full steam'. The disk c0t8d0 does seem to be the most demand as that is where the informix DB is. Does this look as clean cut as a disk i/o bottleneck?
Many thanks for your input so far.

[Tim Sanko]

Hi,

My rule of thumb is 2x on swap 1GB is too small only if you know (from swapinfo) that you are in danger of running out of swap.

Here is my suggestion on kernel changes.

I think your maxdsize is to small. I think that you need to change your maxuprc na d nproc.

maxuprc ((NPROC*9)/4)
nproc ((MAXUSERS*10)+64)
maxdsiz(16384 current) 32768
maxtsiz(16384 current) 32768
timeslice (1 currently) 4

Is this also a hardware issue?
Tim

[Patrick Wallek]

Do you have glance installed on this machine? If not, you can install a trial version from the applications CDs. This will help you in diagnosing what is happening.

Also, do you happen to have 'sar -d' output from the times the machine is feeling slow?

[steve_586]

I will give the kernel changes a try next week.
I have never used Glance but will give it a go and see what it comes up with. Attatched are sar -d for today. Thanks for your continued input.

[Stuart Abramson_2]

Your avserv and avwait times are fairly high.

The %busy on your c0t8d0 are high. It is overloaded compared to the other disks, at least at this time.

You are using 4 internal probably SCSI disks. They are probably slow disks.

However, the disks were slow before! Same disks! So the cpu upgrade shouldn't have made the disks any slower.

I never set the buffer cache min and max equal. I use 5% and 10%. or 2% and 5%.

[Bill Hassell · Accepted solution]

Installing more RAM will not help anything if you weren't using the RAM you already had. To use the extra RAM, you'll need to tell your applications that they can have more (if they are smart enough to do that). Check your kernel parameter dbc_max_pct to see that it is set to 20 (that's 20% or about 150megs). If it was neever changed, then it's 50% and 384 megs may be used for the buffer cache. Memory only helps when:

a. processes are swapping (paging) out. Paging is a very bad feature for performance UNLESS it occurs just occasionally.

b. The buffer cache can be increased (but not more than 200-400 megs max) and the applications can be changed to use more RAM.

swap space has no positive effect on performance! If you have processes that cannot fit into RAM, then something will be swapped out and stop running with a big impact on performance. On the other hand, very interactive processes may only be needed once an hour in which case, the page out/in is very small.

Without good reports from the previous system it will be quite difficult to find what might be slowing things down. Glance will be quite helpful.

[Decio Miname]

I'd start taking a close look at c0t8d0, as Patrick and Stuart mentioned. It appears it is the first clear limit, after improving it you may be able to find others (but then, running faster than now). You could take a look of its contents and try to distribute the most accessed between other disks.

BTW, what model is this disk? The transfer rates and seeks for this scenario are too low even for slow old disks. This could be due to bad internal distribution of LV's inside the disk's sectors.