Re: Performance difference between 2 machines

Etunimi Sukunimi apparently said,on my timestamp of 22/06/2005 8:26 PM:

>
> Okey, I have attached two text files.
>
> There isn't that much difference in write times, but the read times look
> quite alarming.

Very interesting numbers. I'd very much like to see this repeated with twice the size (800Mb) just to confirm the effects of caches in disks/controllers/subsystems.

The results with the smaller IO sizes are very consistent with what I've seen with file system block sizes: HPUX has a sweet spot at 8K which is the default block size for vxfs, while RH has the sweet spot at 4K IO size, the default block size for ext-type file systems. Any IO size below that suffers major degradation. This seems to indicate: make the db block size at least 8K in HPUX and at least 4K in RH and you won't be far wrong.

> Using 16K block size, customer's machine takes a second to complete, but
> our machine takes 0.159 seconds. That is over six times slower.
> That does look conclusive enough for me, but what would you say?

Well, yes: you do have a marked difference there and to me that points to the P4 being much faster than the 9k/800. The key being the deltas between real and sys time. Roughly:
sys time and user time are determined by CPU/memory speed. real time = disk subsystem time + (sys time + user time).

Most of the work is taking place at CPU/memory. Note also how close the real time is to the sys time for reads with both systems: that to me is confirmation the IO is mostly CPU overhead activity. This in turn highlights any CPU speed differences between the two machines.

I'd say also the disk subsystems are very unequal in raw performance, look at the writes. Once you go past the file system sweet spot in write size the determining factor for the RH box is the real time. Not the sys time. Consistent with a fast CPU and a slow writing disk. Whereas in the HP, the major factor is the sys time: that CPU can hardly cope with the disk subsystem speed! Most likely the HP disk subsystem has a huge and fast writeback cache.

There is an anomaly in the 1Mb IO size: HP and other large UNIX systems have optimizations for large IO sizes while RH doesn't until kernel 2.6 (that's why I like to include such a large IO). I suspect this is why the sys time in HPUX is in this case less than the RH one. This could be significant for tuning DBFMBR and FTS.

Reckon you can safely say: make the CPU/memory in the 9k/800 faster if you want to have similar times to your P4. Because the disk subsystem in the HP can cope with faster CPUs better than your P4 disk can.

For the block sizes: I don't see them as a determining factor here at all. Provided they remain as large as (or larger) than the file system ones, you're in diminishing returns land.

Fabrizio, Joel you following this? If so do you agree?

-- 
Cheers
Nuno Souto
in sunny Sydney, Australia
wizofoz2k_at_yahoo.com.au.nospam