Обсуждение: What HW / OS is recommeded
Hi, we are planning to upgrade our servers but deciding on the right configuration seems to be quite difficult. As for the system. About 50 tables, 20M records and growing about 500k-1m per month. The systems mostly loads data from files (perl batch jobs). And generates client files. Jobs generally dont run at the same time, but timely loading/delivery is very important. We also run tomcat with no more than 5-10 concurrent users connecting. (mostly browsing data) We are currently looking at Dell / HP but the questions is - how many processors (2 or 4) - do we gain with 4 cpus if we probably never have a few users connected - what processors are recommended Opteron / Xeon / Itanium - how much memory ? 2GB ? 4GB ? - Disks, i guess we go with Raid5, 15k SCSI - what OS ? Suse / RHE3 / Fedora / - Disk controller ? Currently we run it on a Dell Blade, dual P3 1.4ghz with 1G memory Adding 1GB memory did actually not bring much performance gains. Does anyone have some first hand experience? Can anybody point me to some resources ? Or recommend certain systems? What kind of performance gain can be expected going from a P3 to a higher end processor ? Thanks for any suggestions Alex
I think and please correct me that Postgres loves RAM, the more the better. Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) I use Debian Sarge and im very happy. Perl is very slow, maybe you can use PHP ? Alex wrote: > Hi, > we are planning to upgrade our servers but deciding on the right > configuration seems to be quite difficult. > > As for the system. About 50 tables, 20M records and growing about > 500k-1m per month. > The systems mostly loads data from files (perl batch jobs). And > generates client files. Jobs generally dont run at the same time, but > timely loading/delivery is very important. We also run tomcat with no > more than 5-10 concurrent users connecting. (mostly browsing data) > > We are currently looking at Dell / HP > but the questions is > > - how many processors (2 or 4) > - do we gain with 4 cpus if we probably never have a few users connected > - what processors are recommended Opteron / Xeon / Itanium > - how much memory ? 2GB ? 4GB ? > - Disks, i guess we go with Raid5, 15k SCSI > - what OS ? Suse / RHE3 / Fedora / > - Disk controller ? > > > Currently we run it on a Dell Blade, dual P3 1.4ghz with 1G memory > Adding 1GB memory did actually not bring much performance gains. > > Does anyone have some first hand experience? Can anybody point me to > some resources ? Or recommend certain systems? > What kind of performance gain can be expected going from a P3 to a > higher end processor ? > Thanks for any suggestions > > Alex > > > ---------------------------(end of broadcast)--------------------------- > TIP 5: Have you checked our extensive FAQ? > > http://www.postgresql.org/docs/faqs/FAQ.html -- -------------------------- Canaan Surfing Ltd. Internet Service Providers Ben-Nes Michael - Manager Tel: 972-4-6991122 Cel: 972-52-8555757 Fax: 972-4-6990098 http://www.canaan.net.il --------------------------
Michael Ben-Nes wrote: > I think and please correct me that Postgres loves RAM, the more the better. Certainly for disk-cache. > Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) Raid 10 seems to be the consensus if you have enough disks. See the archives of the performance list for plenty of discussion. > I use Debian Sarge and im very happy. Not sure it makes much difference performance-wise. I'd use whatever flavour of Linux you have the most experience with. > Perl is very slow, maybe you can use PHP ? If your Perl is slower than your PHP, you need to get a better Perl programmer ;-) In any case, if your application is too slow for your database then all the previous tuning is largely irrelevant. -- Richard Huxton Archonet Ltd
Michael Ben-Nes wrote: > I think and please correct me that Postgres loves RAM, the more the better. > > Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) > > I use Debian Sarge and im very happy. > > Perl is very slow, maybe you can use PHP ? I find perl perfectly acceptable. I would appreciate some examples/benchmarks/comparisons? -- Until later, Geoffrey
Hmm... I read that Raid5 is suggested over Raid1. Also HW vendors told us that. Php :-) is not an option and I dont believe Perl is a bottleneck as well. Michael Ben-Nes wrote: > I think and please correct me that Postgres loves RAM, the more the > better. > > Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) > > I use Debian Sarge and im very happy. > > Perl is very slow, maybe you can use PHP ? > > > Alex wrote: > >> Hi, >> we are planning to upgrade our servers but deciding on the right >> configuration seems to be quite difficult. >> >> As for the system. About 50 tables, 20M records and growing about >> 500k-1m per month. >> The systems mostly loads data from files (perl batch jobs). And >> generates client files. Jobs generally dont run at the same time, but >> timely loading/delivery is very important. We also run tomcat with no >> more than 5-10 concurrent users connecting. (mostly browsing data) >> >> We are currently looking at Dell / HP >> but the questions is >> >> - how many processors (2 or 4) >> - do we gain with 4 cpus if we probably never have a few users connected >> - what processors are recommended Opteron / Xeon / Itanium >> - how much memory ? 2GB ? 4GB ? >> - Disks, i guess we go with Raid5, 15k SCSI >> - what OS ? Suse / RHE3 / Fedora / >> - Disk controller ? >> >> >> Currently we run it on a Dell Blade, dual P3 1.4ghz with 1G memory >> Adding 1GB memory did actually not bring much performance gains. >> >> Does anyone have some first hand experience? Can anybody point me to >> some resources ? Or recommend certain systems? >> What kind of performance gain can be expected going from a P3 to a >> higher end processor ? >> Thanks for any suggestions >> >> Alex >> >> >> ---------------------------(end of broadcast)--------------------------- >> TIP 5: Have you checked our extensive FAQ? >> >> http://www.postgresql.org/docs/faqs/FAQ.html > > >
On Thu, 2004-12-16 at 06:39, Michael Ben-Nes wrote: > I think and please correct me that Postgres loves RAM, the more the better. > > Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) With battery backed cache and a large array, RAID 5 is quite fast, even with writes. Plus with a lot of drives in a mostly read environment, it's quite likely that each read will hit a different drive so that many parallel requests can be handled quite well. The general rule I use is 6 or fewer drives will do better in RAID 1+0, 7 or more will tend to do better with RAID 5. > Perl is very slow, maybe you can use PHP ? While mod_perl and its relations have never been fast running under apache in comparison to PHP, it's no slouch, paying mostly in startup time, not run time. For complex apps, the startup time difference becomes noise compared to the run time, so it's no big advantage to PHP. I really like PHP by the way. But Perl is pretty nice too. Run the Unix OS you're most comfortable with, knowing that PostgreSQL gets lots of testing on the free unixes more so than on the commercial ones. Give it a machine with plenty of RAM and a fast I/O subsystem, and two CPUS and you'll get good performance. If your needs exceed the performance of one of these machines, you're probably better off going to a pgpool / slony cluster than trying to build a bigger machine.
Alex wrote: > Hmm... > I read that Raid5 is suggested over Raid1. Also HW vendors told us that. > Php :-) is not an option and I dont believe Perl is a bottleneck as well. Why would your HW vendor be stipulating the software you use? -- Until later, Geoffrey
Quoting Scott Marlowe <smarlowe@g2switchworks.com>: > On Thu, 2004-12-16 at 06:39, Michael Ben-Nes wrote: > > I think and please correct me that Postgres loves RAM, the more the > better. > > > > Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) > > With battery backed cache and a large array, RAID 5 is quite fast, even > with writes. Plus with a lot of drives in a mostly read environment, > it's quite likely that each read will hit a different drive so that many > parallel requests can be handled quite well. The general rule I use is > 6 or fewer drives will do better in RAID 1+0, 7 or more will tend to do > better with RAID 5. > > > Perl is very slow, maybe you can use PHP ? > > While mod_perl and its relations have never been fast running under > apache in comparison to PHP, it's no slouch, paying mostly in startup > time, not run time. For complex apps, the startup time difference > becomes noise compared to the run time, so it's no big advantage to > PHP. I really like PHP by the way. But Perl is pretty nice too. I run apache2, ssl, mod_perl and php. I have yet to hear complaints from my perl or php programmer. Without have another PHP vs. Perl "thing" lets all agree that they are both pretty nice :) > Run the Unix OS you're most comfortable with, knowing that PostgreSQL > gets lots of testing on the free unixes more so than on the commercial > ones. Give it a machine with plenty of RAM and a fast I/O subsystem, > and two CPUS and you'll get good performance. If your needs exceed the > performance of one of these machines, you're probably better off going > to a pgpool / slony cluster than trying to build a bigger machine. I'm not sure I heard any mention of filesystems but I've been moving all my EXT3 filesystems to XFS. Some other journaling filesystem that you might want to look into are JFS and ReiserFS. -- Keith C. Perry, MS E.E. Director of Networks & Applications VCSN, Inc. http://vcsn.com ____________________________________ This email account is being host by: VCSN, Inc : http://vcsn.com
The comment about HW vendor was regarding Raid configuration not the software. Geoffrey wrote: > Alex wrote: > >> Hmm... >> I read that Raid5 is suggested over Raid1. Also HW vendors told us that. >> Php :-) is not an option and I dont believe Perl is a bottleneck as >> well. > > > Why would your HW vendor be stipulating the software you use? >
We use perl for the heavy batch jobs, the web interface is written using JSP / applets. If we would change these then it would be Java or C. But all the heavy stuff is handled by Stored Procedures so I dont see a real need for a change. I actually am more interested to hear if there are an recommended systems or setups. Also with regard to 2/4 CPUs or 32/64 bit etc. Scott Marlowe wrote: >On Thu, 2004-12-16 at 06:39, Michael Ben-Nes wrote: > > >>I think and please correct me that Postgres loves RAM, the more the better. >> >>Any way RAID5 is awful with writing, go with RAID1 ( mirroring ) >> >> > >With battery backed cache and a large array, RAID 5 is quite fast, even >with writes. Plus with a lot of drives in a mostly read environment, >it's quite likely that each read will hit a different drive so that many >parallel requests can be handled quite well. The general rule I use is >6 or fewer drives will do better in RAID 1+0, 7 or more will tend to do >better with RAID 5. > > > >>Perl is very slow, maybe you can use PHP ? >> >> > >While mod_perl and its relations have never been fast running under >apache in comparison to PHP, it's no slouch, paying mostly in startup >time, not run time. For complex apps, the startup time difference >becomes noise compared to the run time, so it's no big advantage to >PHP. I really like PHP by the way. But Perl is pretty nice too. > >Run the Unix OS you're most comfortable with, knowing that PostgreSQL >gets lots of testing on the free unixes more so than on the commercial >ones. Give it a machine with plenty of RAM and a fast I/O subsystem, >and two CPUS and you'll get good performance. If your needs exceed the >performance of one of these machines, you're probably better off going >to a pgpool / slony cluster than trying to build a bigger machine. > > > >
Alex wrote: > The comment about HW vendor was regarding Raid configuration not the > software. My apologies, misread your post. > > > Geoffrey wrote: > >> Alex wrote: >> >>> Hmm... >>> I read that Raid5 is suggested over Raid1. Also HW vendors told us that. >>> Php :-) is not an option and I dont believe Perl is a bottleneck as >>> well. >> >> >> >> Why would your HW vendor be stipulating the software you use? -- Until later, Geoffrey
On Thu, 2004-12-16 at 19:10, Alex wrote: > I actually am more interested to hear if there are an recommended > systems or setups. > Also with regard to 2/4 CPUs or 32/64 bit etc. Sorry to have gotten off on a tangent there. Posts in the last year or so to the -performance mailing list have shown the 64 bit AMD platform to be the fastest X86 based platform around, and having 64 bit hardware is nice for postgresql installations dealing with large data sets. Generally, 2 CPUs is plenty. The fastest storage systems seems to be SAN based, with large local RAID arrays with battery backed cache coming in a close second. Once the RAID array or SAN device has enough drives, putting things like the transaction log elsewhere have little effect. But it's really all about what you're doing with your database. If you're taking huge data sets and running statistical analysis with the plR, you'll need lots of memory as well as plenty of CPU horsepower. If you're handling thousands of simultaneous air line reservations, you'll need lots of drives, and a fair number of CPUs, but probably not so much memory. You may need clusters for one solution, but find one big server is the answer to another problem. There's no one simple answer, because there's no one simple problem. But for most use cases, a dual AMD 64 bit CPU, 4 gigs of ram, and a half a dozen hard drives is a good starting point.
>>>>> "a" == alex <alex@meerkatsoft.com> writes: a> We are currently looking at Dell / HP a> but the questions is a> - how many processors (2 or 4) a> - do we gain with 4 cpus if we probably never have a few users connected a> - what processors are recommended Opteron / Xeon / Itanium a> - how much memory ? 2GB ? 4GB ? a> - Disks, i guess we go with Raid5, 15k SCSI a> - what OS ? Suse / RHE3 / Fedora / a> - Disk controller ? Run, do not walk, from your Dell solution. I've never been able to get "expected" performance from those boxes. They seem to do something to the RAID controllers to make them not work as fast as one would expect the equivalent name-brand part (eg, LSI RAID card or Adaptec RAID card) and similar disk drives. Others have commented on the other aspects. The RAID configuration is something everyone has an opinion for. I personally run a RAID5 on 14 disks, but the new server replacing this Dell box will have RAID10 on 6 disks for data and RAID1 on 2 disks for pg_xlog + system. It is a dual Opteron with 4GB RAM and an Adaptec controller. I only run FreeBSD. See the performance mailing list for a big discussion on vendors just two or three weeks ago. -- =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= Vivek Khera, Ph.D. Khera Communications, Inc. Internet: khera@kciLink.com Rockville, MD +1-301-869-4449 x806 AIM: vivekkhera Y!: vivek_khera http://www.khera.org/~vivek/
On Wed, 22 Dec 2004 12:28:22 -0500, Vivek Khera <khera@kcilink.com> wrote: > >>>>> "a" == alex <alex@meerkatsoft.com> writes: > > a> We are currently looking at Dell / HP > a> but the questions is > > a> - how many processors (2 or 4) > a> - do we gain with 4 cpus if we probably never have a few users connected > a> - what processors are recommended Opteron / Xeon / Itanium > a> - how much memory ? 2GB ? 4GB ? > a> - Disks, i guess we go with Raid5, 15k SCSI > a> - what OS ? Suse / RHE3 / Fedora / > a> - Disk controller ? > > Run, do not walk, from your Dell solution. I've never been able to > get "expected" performance from those boxes. They seem to do > something to the RAID controllers to make them not work as fast as one > would expect the equivalent name-brand part (eg, LSI RAID card or > Adaptec RAID card) and similar disk drives. Hate to burst your bubble, but the RAID controller that Dell ships is an Adaptec OEM. Dell just rebrands them. -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ L. Friedman netllama@gmail.com LlamaLand http://netllama.linux-sxs.org
On Wed, 2004-12-22 at 11:41, Lonni J Friedman wrote: > On Wed, 22 Dec 2004 12:28:22 -0500, Vivek Khera <khera@kcilink.com> wrote: > > >>>>> "a" == alex <alex@meerkatsoft.com> writes: > > > > a> We are currently looking at Dell / HP > > a> but the questions is > > > > a> - how many processors (2 or 4) > > a> - do we gain with 4 cpus if we probably never have a few users connected > > a> - what processors are recommended Opteron / Xeon / Itanium > > a> - how much memory ? 2GB ? 4GB ? > > a> - Disks, i guess we go with Raid5, 15k SCSI > > a> - what OS ? Suse / RHE3 / Fedora / > > a> - Disk controller ? > > > > Run, do not walk, from your Dell solution. I've never been able to > > get "expected" performance from those boxes. They seem to do > > something to the RAID controllers to make them not work as fast as one > > would expect the equivalent name-brand part (eg, LSI RAID card or > > Adaptec RAID card) and similar disk drives. > > Hate to burst your bubble, but the RAID controller that Dell ships is > an Adaptec OEM. Dell just rebrands them. I've use the Dell PERC 4DC and had VERY good performance from it. IT's the late model U320 LSI MegaRAID and runs great. I do remember that the 2650 and few other Dells had the serverworks chipset in them that caused a lot of context switches in heavy parallel load in a discussion on the performance list. We weren't running heavy parallel, just a report server with a dozen or so users, so it wasn't an issue for us.
On Dec 22, 2004, at 12:41 PM, Lonni J Friedman wrote: >> get "expected" performance from those boxes. They seem to do >> something to the RAID controllers to make them not work as fast as one >> would expect the equivalent name-brand part (eg, LSI RAID card or >> Adaptec RAID card) and similar disk drives. > > Hate to burst your bubble, but the RAID controller that Dell ships is > an Adaptec OEM. Dell just rebrands them. They also rebrand LSI. I'm saying that the versions they sell don't seem to perform as well as the manufacturer branded ones from what I've compared with other folks. Vivek Khera, Ph.D. +1-301-869-4449 x806
On Dec 22, 2004, at 1:09 PM, Scott Marlowe wrote: > I've use the Dell PERC 4DC and had VERY good performance from it. IT's > the late model U320 LSI MegaRAID and runs great. I do remember that > the > 2650 and few other Dells had the serverworks chipset in them that > caused > a lot of context switches in heavy parallel load in a discussion on the > performance list. We weren't running heavy parallel, just a report > server with a dozen or so users, so it wasn't an issue for us. > Whatever the ultimate cause, I can't say. But I can say that I'm not planning to buy any more Dell servers for my database needs. Both my 2450 and 2650 basically suck at I/O load with any kind of concurrent access. I might as well be running IDE :-(
Lonni J Friedman wrote: > On Wed, 22 Dec 2004 12:28:22 -0500, Vivek Khera <khera@kcilink.com> wrote: > >>>>>>>"a" == alex <alex@meerkatsoft.com> writes: >> >>a> We are currently looking at Dell / HP >>a> but the questions is >> >>a> - how many processors (2 or 4) >>a> - do we gain with 4 cpus if we probably never have a few users connected >>a> - what processors are recommended Opteron / Xeon / Itanium >>a> - how much memory ? 2GB ? 4GB ? >>a> - Disks, i guess we go with Raid5, 15k SCSI >>a> - what OS ? Suse / RHE3 / Fedora / >>a> - Disk controller ? >> >>Run, do not walk, from your Dell solution. I've never been able to >>get "expected" performance from those boxes. They seem to do >>something to the RAID controllers to make them not work as fast as one >>would expect the equivalent name-brand part (eg, LSI RAID card or >>Adaptec RAID card) and similar disk drives. > > > Hate to burst your bubble, but the RAID controller that Dell ships is > an Adaptec OEM. Dell just rebrands them. Dell is known for having their own 'version' of hardware made, specifically to reduce the cost of the machines. I don't know that this is the case with this adapter, but I would not ever purchase another dell machine. -- Until later, Geoffrey