Обсуждение: benchmarks with pgbench
Hi Guys, I''ve been running pgbench tests for a while. I have one server set up to run pgbench tpc tests (7.x). A new server that I just configured with SUSE and 8.0.0 just gets killed even though it has the same memory 8 GB and it's a x86-64 box. The other one is Solaris full 64-bit. That seems reasonable, but we thought postgres would run better on a linux box than solaris. Also, the x86-64 box does much worse than a 32-bit linux box with mandrake and a lot less RAM. I've restarted postgres with many different postgresql.conf configs and nothing seems to make much of a difference to pgbench. Has anyone else experienced slower performance on 64-bit linux as compared to 32-bit linux? D.J.
On Mon, 2005-01-24 at 10:48, Kavan, Dan (IMS) wrote: > Hi Guys, > I''ve been running pgbench tests for a while. I have one server set up > to run pgbench tpc tests (7.x). A new server that I just configured > with SUSE and 8.0.0 just gets killed even though it has the same memory > 8 GB and it's a x86-64 box. The other one is Solaris full 64-bit. That > seems reasonable, but we thought postgres would run better on a linux > box than solaris. Also, the x86-64 box does much worse than a 32-bit > linux box with mandrake and a lot less RAM. I've restarted postgres > with many different postgresql.conf configs and nothing seems to make > much of a difference to pgbench. Has anyone else experienced slower > performance on 64-bit linux as compared to 32-bit linux? pgbench is notorious for providing poor measure of a database's performance under real world load. Are you sure your Solaris and Linux boxes are both running on SCSI hard drives (IDE drives are well known for not obeying fsync() calls, but simply saying "yep, synced that data" when in fact they haven't. So, if you Linux box is set to both fsync properly AND is writing access time to each file, it may be quite a bit slower than a Solaris box if that box is writing to IDE drives, has fsync turned off, and / or has access time writing disabled.
Yes, they are both running on the same hardware - NAS. On Mon, 2005-01-24 at 10:48, Kavan, Dan (IMS) wrote: > Hi Guys, > I''ve been running pgbench tests for a while. I have one server set > up to run pgbench tpc tests (7.x). A new server that I just > configured with SUSE and 8.0.0 just gets killed even though it has the > same memory 8 GB and it's a x86-64 box. The other one is Solaris full > 64-bit. That seems reasonable, but we thought postgres would run > better on a linux box than solaris. Also, the x86-64 box does much > worse than a 32-bit linux box with mandrake and a lot less RAM. I've > restarted postgres with many different postgresql.conf configs and > nothing seems to make much of a difference to pgbench. Has anyone > else experienced slower performance on 64-bit linux as compared to > 32-bit linux? pgbench is notorious for providing poor measure of a database's performance under real world load. Are you sure your Solaris and Linux boxes are both running on SCSI hard drives (IDE drives are well known for not obeying fsync() calls, but simply saying "yep, synced that data" when in fact they haven't. So, if you Linux box is set to both fsync properly AND is writing access time to each file, it may be quite a bit slower than a Solaris box if that box is writing to IDE drives, has fsync turned off, and / or has access time writing disabled.
I hate to admit this publically, but I've been reading my results backwards. I was getting 100 tps on Solaris - postgres 64 bit and 300 tps on SUSE postgres both x86-64. So, 300 is better than 100 right? I was reading it backwards. I was thinking 300 was the actual speed to process a certain amount of transactions, but actually the x86-64 system is performing better than all, not worse. ~Dj Yes, they are both running on the same hardware - NAS. On Mon, 2005-01-24 at 10:48, Kavan, Dan (IMS) wrote: > Hi Guys, > I''ve been running pgbench tests for a while. I have one server set > up to run pgbench tpc tests (7.x). A new server that I just > configured with SUSE and 8.0.0 just gets killed even though it has the > same memory 8 GB and it's a x86-64 box. The other one is Solaris full > 64-bit. That seems reasonable, but we thought postgres would run > better on a linux box than solaris. Also, the x86-64 box does much > worse than a 32-bit linux box with mandrake and a lot less RAM. I've > restarted postgres with many different postgresql.conf configs and > nothing seems to make much of a difference to pgbench. Has anyone > else experienced slower performance on 64-bit linux as compared to > 32-bit linux? pgbench is notorious for providing poor measure of a database's performance under real world load. Are you sure your Solaris and Linux boxes are both running on SCSI hard drives (IDE drives are well known for not obeying fsync() calls, but simply saying "yep, synced that data" when in fact they haven't. So, if you Linux box is set to both fsync properly AND is writing access time to each file, it may be quite a bit slower than a Solaris box if that box is writing to IDE drives, has fsync turned off, and / or has access time writing disabled. ---------------------------(end of broadcast)--------------------------- TIP 6: Have you searched our list archives? http://archives.postgresql.org
On Mon, 2005-01-24 at 13:19, Kavan, Dan (IMS) wrote: > I hate to admit this publically, but I've been reading my results > backwards. > > I was getting 100 tps on Solaris - postgres 64 bit and 300 tps on SUSE > postgres both x86-64. > So, 300 is better than 100 right? I was reading it backwards. > I was thinking 300 was the actual speed to process a certain amount of > transactions, but actually the x86-64 system is performing better than > all, not worse. > > ~Dj > > > > Yes, they are both running on the same hardware - NAS. Yep, 300 is better than 100. Glad to hear it.