From: Tarasyuk Nik (NTarasyuk@snowyhydro.com.au)
Date: Tue Apr 23 2002 - 02:46:10 EDT
Hi
Sorry for the late summary, but the issue is still dragging on.
Thanks to Selden E Ball Jr, alan@nabeth.cxo.cpqcorp.net, Werner Hahling, Murtaza Talawala,
Shaukat Riaz, Udo de Boer and others for excellent insights into the subject.
We have done more testing and excluded many unknowns from our equation,
e.g, there is no problem with disk I/O and network I/O.
We have lots of memory, I put in twice more than it was required.
There is enough space for processes and UCB, disk cach hit rate is 99%.
On another hand, CPU utilization is still 100%, and system time is quite high.
It looks like this
tty fd0 rz8 dk2 dk3 cpu
tin tout bps tps bps tps bps tps bps tps us ni sy id
0 222 0 0 104 14 0 0 0 0 44 0 56 0
0 478 0 0 92 5 0 0 0 0 42 0 58 0
0 535 0 0 63 5 0 0 0 0 41 0 59 0
and this
Virtual Memory Statistics: (pagesize = 8192)
procs memory pages intr cpu
r w u act free wire fault cow pin pout in sy cs us sy id iowait
8 631 23 43K 9419 9689 1952 201 229 0 427 38K 3K 36 64 0 0
8 640 23 43K 8978 9816 3206 324 283 0 479 32K 4K 48 52 0 0
5 651 23 44K 8267 9901 4307 563 611 0 293 23K 2K 50 50 0 0
4 653 23 44K 8174 9912 1261 85 86 0 295 25K 2K 43 57 0 0
and this
input (tu0) output input (Total) output
packets errs packets errs colls packets errs packets errs colls
754 0 755 5 26 10170 0 10171 5 26
8 0 8 0 0 159 0 159 0 0
22 0 22 0 0 195 0 195 0 0
8 0 7 0 0 163 0 162 0 0
110 0 124 0 0 456 0 470 0 0
Performance management manual says that CPU system time over 50% is not good.
So, probably, our application is cauzing it.
We see 2 ways out:
1. buy better CPU
2. tune application, and system configuration, like IPC to reduce system time.
I have to note here, that to tune system parameters like IPC,
performance manual is not enough. You can pick up a lot of undestanding from there,
but in many cases it stops short at suggesting concrete parameters values.
For example it suggests to increase some value which has 16000 default.
All right, by how much? Twice? Three times? Ten times?
No indication of that.
BTW I did run sys_check and followed its advices. Didn't help.
So, meanwhile, we kicked it back to the application developers,
and looking for cash to buy new boxes.
Best Regards
Nik Tarasyuk
Software Engineer
Snowy Hydro
Australia
-----Original Message-----
I am looking to increase performance of Alphaserevr 1000A 4/266 box running 4.0f.
It has everything in it:
quite heavy math, high network untilization (X traffic and other),
also it has special PCI interface card, and application that works with it takes around 40% of CPU,
it's disk access intensive, too.
Sorry for the lack of details and logs.
We were given a very narrow window on the live system and performance was last on our minds.
The issue was discovered during last minutes, so that's all I have, info from "top".
The box has 512Mb memory with only 320Mb utilized, so the memory wouldn't be an issue.
At the moment I am thinking of the following:
1. Will use faster harddrive (10000rpm, LVD support) with SYM8952U Wide Ultra2 instead of internal SCSI controller.
This is easy to do for us, we have everything available except wide SCSI cable.
This would maximize disk access.
2. At the moment that PCI interface card is plugged into secondary PCI bus.
I wonder if replugging it into primary PCI bus would help somehow (it's bridged architecture),
so I'll do it anyway, it shouldn't hurt.
But we have general feeling that the task is simply over the head of 4/266 CPU we have
and it needs to be upgraded.
Unfortunately, I might be able to repeat the run and take good performance tests only in, at least, one week time...
Please, share your thoughts on this. Any info might prove to be helpful.
Thanks
Nik Tarasyuk
Software Engineer
Snowy Hydro
Australia
This archive was generated by hypermail 2.1.7 : Sat Apr 12 2008 - 10:48:39 EDT