Обсуждение: general peformance question
Hi, This a kind of newbie-question. I've been using Postgres for a long time in a low transction environment - and it is great. Now I've got an inquiry for using Postgresql in a heavy-load on-line system. This system must handle something like 20 questionsper sec with a response time at 1/10 sec. Each question will result in approx 5-6 reads and a couple of updates. Anybody have a feeling if this is realistic on a Intelbased Linux server with Postgresql. Ofcourse I know that this is toolittle info for an exact answer but - as I said - maybe someone can give a hint if it's possible. Maybe someone with heavy-loadcan give an example of what is possible... Regards Conny Thimrén
On Thu, 2003-12-18 at 12:04, Conny Thimren wrote: > Hi, > This a kind of newbie-question. I've been using Postgres for a long time in a low transction environment - and it is great. > > Now I've got an inquiry for using Postgresql in a heavy-load on-line system. This system must handle something like 20questions per sec with a response time at 1/10 sec. Each question will result in approx 5-6 reads and a couple of updates. > Anybody have a feeling if this is realistic on a Intelbased Linux server with Postgresql. Ofcourse I know that this istoo little info for an exact answer but - as I said - maybe someone can give a hint if it's possible. Maybe someone withheavy-load can give an example of what is possible... Ok, is that 20 questions per second (20 in parallel taking 1 second each) or serialized taking 50ms each. Are they simple selects / updates (less than 10 rows in result set, very simple joins) or are they monster 30 table join queries?
On Thu, 18 Dec 2003, Conny Thimren wrote: > Hi, > This a kind of newbie-question. I've been using Postgres for a long time in a low transction environment - and it is great. > > Now I've got an inquiry for using Postgresql in a heavy-load on-line system. This system must handle something like 20questions per sec with a response time at 1/10 sec. Each question will result in approx 5-6 reads and a couple of updates. > Anybody have a feeling if this is realistic on a Intelbased Linux server with Postgresql. Ofcourse I know that this istoo little info for an exact answer but - as I said - maybe someone can give a hint if it's possible. Maybe someone withheavy-load can give an example of what is possible... That really depends on how heavy each query is, so it's hard to say from what little you've given us. If you are doing simple banking style transactions, then you can easily handle this load, if you are talking a simple shopping cart, ditto, if, however, you are talking about queries that run 4 or 5 tables with millions of rows againts each other, you're gonna have to test it yourself. With the autovacuum daemon running, I ran a test overnight of pgbench (more for general purpose burn in than anything else) pgbench -i -s 100 pgbench -c 50 -t 250000 that's 10 million transactions, and it took just over twelve hours to complete at 220+ transactions per second. so, for financials, you're likely to find it easy to meet your target. But as the tables get bigger / more complex / more interconnected you'll see a drop in performance.