On 5/13/09 3:22 AM, "Dimitri" <dimitrik.fr@xxxxxxxxx> wrote: > Hi Scott, > > On 5/12/09, Scott Carey <scott@xxxxxxxxxxxxxxxxx> wrote: >> Although nobody wants to support it, he should try the patch that Jignesh K. >> Shah (from Sun) proposed that makes ProcArrayLock lighter-weight. If it >> makes 32 cores much faster, then we have a smoking gun. >> >> Although everyone here is talking about this as an 'unoptimal' solution, the >> fact is there is no evidence that a connection pooler will fix the >> scalability from 16 > 32 cores. >> Certainly a connection pooler will help most results, but it may not fix the >> scalability problem. >> >> A question for Dimitri: >> What is the scalability from 16 > 32 cores at the 'peak' load that occurs >> near 2x the CPU count? Is it also poor? If this is also poor, IMO the >> community here should not be complaining about this unopimal case -- a >> connection pooler at that stage does little and prepared statements will >> increase throughput but not likely alter scalability. > > I'm attaching a small graph showing a TPS level on PG 8.4 depending on > number of cores (X-axis is a number of concurrent users, Y-axis is the > TPS number). As you may see TPS increase is near linear while moving > from 8 to 16 cores, while on 32cores even it's growing slightly > differently, what is unclear is why TPS level is staying limited to > 11.000 TPS on 32cores. And it's pure read-only workload. > Interesting. What hardware is this, btw? Looks like the 32 core system probably has 2x the CPU and a bit less interconnect efficiency versus the 16 core one (which would be typical). Is the 16 core case the same, but with fewer cores per processor active? Or fewer processors total? Understanding the scaling difference may require a better understanding of the other differences besides core count. >> >> If that result scales, then the short term answer is a connection pooler. >> >> In the tests that Jingesh ran -- making the ProcArrayLock faster helped the >> case where connections = 2x the CPU core count quite a bit. >> >> The thread about the CPU scalability is "Proposal of tunable fix for >> scalability of 8.4", originally posted by "Jignesh K. Shah" >> <J.K.Shah@xxxxxxx>, March 11 2009. >> >> It would be very useful to see results of this benchmark with: >> 1. A Connection Pooler > > will not help, as each client is *not* disconnecting/reconnecting > during the test, as well PG is keeping well even 256 users. And TPS > limit is reached already on 64 users, don't think pooler will help > here. > Actually, it might help a little. Postgres has a flaw that makes backends block on a lock briefly based on the number of total backends -- active or completely passive. Your tool has some (very small) user-side delay and a connection pooler would probably allow 64 of your users to efficiently 'fit' in 48 or so connection pooler slots. It is not about connecting and disconnecting in this case, its about minimizing Postgres' process count. If this does help, it would hint at certain bottlenecks. If it doesn't it would point elsewhere (and quiet some critics). However, its unrealistic for any process-per-connection system to have less backends than about 2x the core count -- else any waiting on I/O or network will just starve CPU. So this would just be done for research, not a real answer to making it scale better. For those who say "but, what if its I/O bound! You don't need more backends then!": Well you don't need more CPU either if you're I/O bound. By definition, CPU scaling tests imply the I/O can keep up. >> 2. Jignesh's patch > > I've already tested it and it did not help in my case because the real > problem is elsewhere.. (however, I did not test it yet with my latest > config params) > Great to hear that! -- That means this case is probably not ProcArrayLock. If its Solaris, could we get: 1. What is the CPU stats when it is in the inefficient state near 64 or 128 concurrent users (vmstat, etc. I'm interested in CPU in user/system/idle/wait time, and context switches/sec mostly). 2. A Dtrace probe on the postgres locks -- we might be able to identify something here. The results here would be useful -- if its an expected condition in the planner or parser, it would be useful confirmation. If its something unexpected and easy to fix -- it might be changed relatively soon. If its not easy to detect, it could be many other things -- but the process above at least rules some things out and better characterizes the state. >> 3. Prepared statements >> > > yes, I'm preparing this test. > >> #3 is important, because prepared statements are ideal for queries that >> perform well with low statistics_targets, and not ideal for those that >> require high statistics targets. Realistically, an app won't have more than >> a couple dozen statement forms to prepare. Setting the default statistics >> target to 5 is just a way to make some other query perform bad. > > Agree, but as you may have a different statistic target *per* table it > should not be a problem. What is sure - all time spent on parse and > planner will be removed here, and the final time should be a pure > execution. > I'm definitely interested here because although pure execution will certainly be faster, it may not scale any better. > Rgds, > -Dimitri > -- Sent via pgsql-performance mailing list (pgsql-performance@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
