Thread: intel vs amd benchmark for pg server part 2
Because of a lot of unwarranted criticizms of my previous rigorous and outstanding :) benchmark and because I am lucky to come across a bargain maxtor drive I have done the test again. It put the same hard drive in both amd 1.33GHz and celeron 566MHz machine with exactly the same controller that comes with the drive. It's promise PDC20269 chipset and I have to run 2.4.19-pre7-ac2 to detect this controller and linux detects it as udma_133 drive. I point PGDATA to the directory on this hard drive on both machines. On amd # hdparm -tT /dev/hde /dev/hde: Timing buffer-cache reads: 128 MB in 0.78 seconds =164.10 MB/sec Timing buffered disk reads: 64 MB in 1.59 seconds = 40.25 MB/sec On celeron the first number is around 100 MB/sec and the second number is about the same. The drive is not in that system anymore and I didn't write the numbers down. The minimun time to vacuum on celeron I got was 61 seconds and whereas on amd it is 59 seconds. So it looks like vacuuming scales very linearly with hdparm results. And yes, with on board ide controller, AMD/sis box hdparm bottleneck is around 24MB/sec whereas corresponding number for intel box is 30MB/sec. The moral I get from this benchmark is that AMD is certainly not much better preformance/price wise at least not for a database server In fact for various reasons I am going to go with an intel box. Another gripe I have is that vacuum process does not eat up 100% of cpu. In the beginning it peaks around 80% and at the end it is stuck around 20%. Whenever I have a long running process and it is not eating up 100% of cpu I feel I am not getting my money's worth for the cpu. I wonder why vacuum process is not more parallelized if at all. I can imagine manually vacuuming many tables in parallel and it might eat up all cpu and I wonder whether it might finish quicker.
On Fri, Apr 26, 2002 at 10:52:35PM -0400, postgres@vrane.com wrote: > Another gripe I have is that vacuum process does not eat up 100% > of cpu. In the beginning it peaks around 80% and at the end > it is stuck around 20%. > > Whenever I have a long running process and > it is not eating up 100% of cpu I feel I am not getting my money's > worth for the cpu. I wonder why vacuum process is not more parallelized > if at all. I can imagine manually vacuuming many tables in parallel > and it might eat up all cpu and I wonder whether it might finish quicker. I think you must not have very much experience with databases. They are totally limited by the disk (storage) subsystem. You know, the part of your machine that has to wait 10 or more milliseconds for a little metal arm to swing into position. Very retro! I'd be pissed off if vacuum was taking that much of my CPU time. Usually it wails away on the disks (8 15000RPM monsters) and the CPUs twiddle their thumbs. As for IDE controllers: they are all fifth-rate crap. Get a SCSI controller. -jwb
Erm. I think you've missed the point slightly. All that CPU power is so that you can watch movies and play MP3s whilst waiting for VACUUM to finish[1]. Be sure to have your movie files on another disk and I/O channel, so that frames don't get lost - but of course ;). Note: if you really don't need the CPU power, what you could do is underclock your CPU, that way you can run cooler and more reliably (maybe even go fanless) or you can get someone else to do it for you and buy a Celeron. Hope that helps, Link. [1] When you are vacuuming really large databases you might even want to crack crypto whilst you're at it :). Or do some ray tracing, or machinima/animated films. Of course with the advances of Postgresql 7.2 these times of repose and reflection will become briefer, rarer and more precious. How we will miss glimpses of The Boss yelling and hammering on the locked server room door whilst the vacuum is going on. But take heart, for the Benevolent Developers in their great wisdom and grace hath provided VACUUM FULL, and our disk buffers overfloweth with joy. At 10:52 PM 4/26/02 -0400, postgres@vrane.com wrote: >The moral I get from this benchmark is that AMD is certainly >not much better preformance/price wise at least not for >a database server > >In fact for various reasons I am going to go with an intel >box. > >Another gripe I have is that vacuum process does not eat up 100% >of cpu. In the beginning it peaks around 80% and at the end >it is stuck around 20%. > >Whenever I have a long running process and >it is not eating up 100% of cpu I feel I am not getting my money's >worth for the cpu. I wonder why vacuum process is not more parallelized >if at all. I can imagine manually vacuuming many tables in parallel >and it might eat up all cpu and I wonder whether it might finish quicker.
On Fri, 26 Apr 2002 postgres@vrane.com wrote: > Another gripe I have is that vacuum process does not eat up 100% > of cpu. In the beginning it peaks around 80% and at the end > it is stuck around 20%. That's because your disk subsystem is too slow for the machine. Put in a disk subsystem that doesn't slow down the machine, and you'll use all your CPU. Then you can complain about not using all your disk I/O capacity. Performance bottlenecks never go away. You can only move them around. cjs -- Curt Sampson <cjs@cynic.net> +81 90 7737 2974 http://www.netbsd.org Don't you know, in this new Dark Age, we're all light. --XTC
Curt Sampson wrote: >On Fri, 26 Apr 2002 postgres@vrane.com wrote: > >>Another gripe I have is that vacuum process does not eat up 100% >>of cpu. In the beginning it peaks around 80% and at the end >>it is stuck around 20%. >> > >That's because your disk subsystem is too slow for the machine. >Put in a disk subsystem that doesn't slow down the machine, and >you'll use all your CPU. > >Then you can complain about not using all your disk I/O capacity. > >Performance bottlenecks never go away. You can only move them around. > Thats why we all call it "chasing the brass ring" :) > >cjs >
Subject: Re: [GENERAL] intel vs amd benchmark for pg server part 2 From: Vic Cekvenich <vic@basebeans.com> === Or a disk caching controler (like Mylex or Penging Computing) Any DB is IO bound. Michael Loftis wrote: > > > Curt Sampson wrote: > >> On Fri, 26 Apr 2002 postgres@vrane.com wrote: >> >>> Another gripe I have is that vacuum process does not eat up 100% >>> of cpu. In the beginning it peaks around 80% and at the end >>> it is stuck around 20%. >>> >> >> That's because your disk subsystem is too slow for the machine. >> Put in a disk subsystem that doesn't slow down the machine, and >> you'll use all your CPU. >> >> Then you can complain about not using all your disk I/O capacity. >> >> Performance bottlenecks never go away. You can only move them around. >> > Thats why we all call it "chasing the brass ring" :) > >> >> cjs >> > > > > ---------------------------(end of broadcast)--------------------------- > TIP 2: you can get off all lists at once with the unregister command > (send "unregister YourEmailAddressHere" to majordomo@postgresql.org)