Thread: Clarification on some settings
Hello, I've been having some performance issues with a DB I use. I'm trying to come up with some performance recommendations to send to the "adminstrator". Hardware: CPU0: Pentium III (Coppermine) 1000MHz (256k cache) CPU1: Pentium III (Coppermine) 1000MHz (256k cache) Memory: 3863468 kB (4 GB) OS: Red Hat Linux release 7.2 (Enigma) Kernel: 2.4.9-31smp I/O I believe is a 3-disk raid 5. /proc/sys/kernel/shmmax and /proc/sys/kernel/shmall were set to 2G Postgres version: 7.3.4 I know its a bit dated, and upgrades are planned, but several months out. Load average seems to hover between 1.0 and 5.0-ish during peak hours. CPU seems to be the limiting factor but I'm not positive (cpu utilization seems to be 40-50%). We have 2 of those set up as the back end to 3 web-servers each... supposedly load-balanced, but one of the 2 dbs consistently has higher load. We have a home-grown replication system that keeps them in sync with each other... peer to peer (master/master). The DB schema is, well to put it nicely... not exactly normalized. No constraints to speak of except for the requisite not-nulls on the primary keys (many of which are compound). Keys are mostly varchar(256) fields. Ok for what I'm uncertain of... shared_buffers: According to http://www.varlena.com/varlena/GeneralBits/Tidbits/perf.html Its more of a staging area and more isn't necessarily better. That psql relies on the OS to cache data for later use. But according to http://www.ca.postgresql.org/docs/momjian/hw_performance/node3.html its where psql caches previous data for queries because the OS cache is slower, and should be as big as possible without causing swap. Those seem to be conflicting statements. In our case, the "administrator" kept increasing this until performance seemed to increase, which means its now 250000 (x 8k is 2G). Is this just a staging area for data waiting to move to the OS cache, or is this really the area that psql caches its data? effective_cache_size: Again, according to the Varlena guide this tells psql how much system memory is available for it to do its work in. until recently, this was set at the default value of 1000. It was just recently increased to 180000 (1.5G) according to http://www.varlena.com/varlena/GeneralBits/Tidbits/annotated_conf_e.html it should be about 25% of memory? Finally sort_mem: Was until recently left at the default of 1000. Is now 16000. Increasing the effective cache and sort mem didn't seem to make much of a difference. I'm guessing the eff cache was probably raised a bit too much, and shared_buffers is way to high. What can I do to help determine what the proper settings should be and/or look at other possible choke points. What should I look for in iostat, mpstat, or vmstat as red flags that cpu, memory, or i/o bound? DB maintenance wise, I don't believe they were running vacuum full until I told them a few months ago that regular vacuum analyze no longer cleans out dead tuples. Now normal vac is run daily, vac full weekly (supposedly). How can I tell from the output of vacuum if the vac fulls aren't being done, or not done often enough? Or from the system tables, what can I read? Is there anywhere else I can look for possible clues? I have access to the DB super-user, but not the system root/user. Thank you for your time. Please let me know any help or suggestions you may have. Unfortunately upgrading postgres, OS, kernel, or re-writing schema is most likely not an option.
Doug Y wrote: > Hello, > I've been having some performance issues with a DB I use. I'm trying > to come up with some performance recommendations to send to the > "adminstrator". > > Hardware: > CPU0: Pentium III (Coppermine) 1000MHz (256k cache) > CPU1: Pentium III (Coppermine) 1000MHz (256k cache) > Memory: 3863468 kB (4 GB) > OS: Red Hat Linux release 7.2 (Enigma) > Kernel: 2.4.9-31smp > I/O I believe is a 3-disk raid 5. > > /proc/sys/kernel/shmmax and /proc/sys/kernel/shmall were set to 2G > > Postgres version: 7.3.4 > The DB schema is, well to put it nicely... not exactly normalized. No > constraints to speak of except for the requisite not-nulls on the > primary keys (many of which are compound). Keys are mostly varchar(256) > fields. > > Ok for what I'm uncertain of... > shared_buffers: > According to http://www.varlena.com/varlena/GeneralBits/Tidbits/perf.html > Its more of a staging area and more isn't necessarily better. That psql > relies on the OS to cache data for later use. > But according to > http://www.ca.postgresql.org/docs/momjian/hw_performance/node3.html its > where psql caches previous data for queries because the OS cache is > slower, and should be as big as possible without causing swap. > Those seem to be conflicting statements. In our case, the > "administrator" kept increasing this until performance seemed to > increase, which means its now 250000 (x 8k is 2G). > Is this just a staging area for data waiting to move to the OS cache, or > is this really the area that psql caches its data? It is the area where postgresql works. It updates data in this area and pushes it to OS cache for disk writes later. By experience, larger does not mean better for this parameter. For multi-Gig RAM machines, the best(on an average for wide variety of load) value found to be around 10000-15000. May be even lower. It is a well known fact that raising this parameter unnecessarily decreases the performance. You indicate that best performance occurred at 250000. This is very very large compared to other people's experience. > > effective_cache_size: > Again, according to the Varlena guide this tells psql how much system > memory is available for it to do its work in. > until recently, this was set at the default value of 1000. It was just > recently increased to 180000 (1.5G) > according to > http://www.varlena.com/varlena/GeneralBits/Tidbits/annotated_conf_e.html > it should be about 25% of memory? No rule of thumb. It is amount of memory OS will dedicate to psotgresql data buffers. Depending uponn what else you run on machine, it could be straight-forward or noodly value to calculate. For a 4GB machine, 1.5GB is quite good but coupled with 2G of shared buffers it could push the machines to swap storm. And swapping shared buffers is a big performance hit. > > Finally sort_mem: > Was until recently left at the default of 1000. Is now 16000. Sort memory is per sort not per query or per connection. So depending upon how many concurrent connections you entertain, it could take quite a chuck of RAM. > > Increasing the effective cache and sort mem didn't seem to make much of > a difference. I'm guessing the eff cache was probably raised a bit too > much, and shared_buffers is way to high. I agree. For shared buffers start with 5000 and increase in batches on 1000. Or set it to a high value and check with ipcs for maximum shared memory usage. If share memory usage peaks at 100MB, you don't need more than say 120MB of buffers. > > What can I do to help determine what the proper settings should be > and/or look at other possible choke points. What should I look for in > iostat, mpstat, or vmstat as red flags that cpu, memory, or i/o bound? Yes. vmstat is usually a lot of help to locate the bottelneck. > DB maintenance wise, I don't believe they were running vacuum full until > I told them a few months ago that regular vacuum analyze no longer > cleans out dead tuples. Now normal vac is run daily, vac full weekly > (supposedly). How can I tell from the output of vacuum if the vac fulls > aren't being done, or not done often enough? Or from the system tables, > what can I read? In 7.4 you can do vacuum full verbose and it will tell you the stats at the end. For 7.3.x, its not there. I suggest you vacuum full database once.(For large database, dumping restoring might work faster. Dump/restore and vacuum full both lock the database exclusively i.e. downtime. So I guess faster the better for you. But there is no tool/guideline to determine which way to go.) > Is there anywhere else I can look for possible clues? I have access to > the DB super-user, but not the system root/user. Other than hardware tuning, find out slow/frequent queries. Use explain analyze to determine why they are so slow. Forgetting to typecast a where clause and using sequential scan could cost you lot more than mistuned postgresql configuration. > Thank you for your time. Please let me know any help or suggestions you > may have. Unfortunately upgrading postgres, OS, kernel, or re-writing > schema is most likely not an option. I hope you can change your queries. HTH Shridhar
On Wed, 2004-05-12 at 05:02, Shridhar Daithankar wrote: > I agree. For shared buffers start with 5000 and increase in batches on 1000. Or > set it to a high value and check with ipcs for maximum shared memory usage. If > share memory usage peaks at 100MB, you don't need more than say 120MB of buffers. If your DB touches more than 100MB worth of buffers over time, shared memory consumption won't peak at 100MB. PG shared buffers are only "recycled" when there are no unused buffers available, so this isn't a really valid way to determine the right shared_buffers setting. -Neil
Note that effective_cache_size is merely a hint to that planner to say "I have this much os buffer cache to use" - it is not actually allocated. It is shared_buffers that will hurt you if it is too high (10000 - 25000 is the usual sweet spot). best wishes Mark Shridhar Daithankar wrote: > >> >> Increasing the effective cache and sort mem didn't seem to make much >> of a difference. I'm guessing the eff cache was probably raised a bit >> too much, and shared_buffers is way to high. > > > I agree. For shared buffers start with 5000 and increase in batches on > 1000. Or set it to a high value and check with ipcs for maximum shared > memory usage. If share memory usage peaks at 100MB, you don't need > more than say 120MB of buffers. > > >
(Sorry if this ends up being a duplicate post, I sent a reply yesterday, but it doesn't appear to have gone through... I think I typo'd the address but never got a bounce.) Hi, Thanks for your initial help. I have some more questions below. At 05:02 AM 5/12/2004, Shridhar Daithankar wrote: >Doug Y wrote: > >>Hello, >> I've been having some performance issues with a DB I use. I'm trying >> to come up with some performance recommendations to send to the "adminstrator". >> >>Ok for what I'm uncertain of... >>shared_buffers: >>According to http://www.varlena.com/varlena/GeneralBits/Tidbits/perf.html >>Its more of a staging area and more isn't necessarily better. That psql >>relies on the OS to cache data for later use. >>But according to >>http://www.ca.postgresql.org/docs/momjian/hw_performance/node3.html its >>where psql caches previous data for queries because the OS cache is >>slower, and should be as big as possible without causing swap. >>Those seem to be conflicting statements. In our case, the "administrator" >>kept increasing this until performance seemed to increase, which means >>its now 250000 (x 8k is 2G). >>Is this just a staging area for data waiting to move to the OS cache, or >>is this really the area that psql caches its data? > >It is the area where postgresql works. It updates data in this area and >pushes it to OS cache for disk writes later. > >By experience, larger does not mean better for this parameter. For >multi-Gig RAM machines, the best(on an average for wide variety of load) >value found to be around 10000-15000. May be even lower. > >It is a well known fact that raising this parameter unnecessarily >decreases the performance. You indicate that best performance occurred at >250000. This is very very large compared to other people's experience. Ok. I think I understand a bit better now. >>effective_cache_size: >>Again, according to the Varlena guide this tells psql how much system >>memory is available for it to do its work in. >>until recently, this was set at the default value of 1000. It was just >>recently increased to 180000 (1.5G) >>according to >>http://www.varlena.com/varlena/GeneralBits/Tidbits/annotated_conf_e.html >>it should be about 25% of memory? > >No rule of thumb. It is amount of memory OS will dedicate to psotgresql >data buffers. Depending uponn what else you run on machine, it could be >straight-forward or noodly value to calculate. For a 4GB machine, 1.5GB is >quite good but coupled with 2G of shared buffers it could push the >machines to swap storm. And swapping shared buffers is a big performance hit. We don't seem to be swapping much: # top 2:21pm up 236 days, 19:12, 1 user, load average: 1.45, 1.09, 1.00 53 processes: 51 sleeping, 2 running, 0 zombie, 0 stopped CPU0 states: 30.3% user, 9.1% system, 0.0% nice, 60.0% idle CPU1 states: 32.0% user, 9.3% system, 0.0% nice, 58.1% idle Mem: 3863468K av, 3845844K used, 17624K free, 2035472K shrd, 198340K buff Swap: 1052248K av, 1092K used, 1051156K free 1465112K cached looks like at some point it did swap a little, but from running vmstat, I can't seem to catch it actively swapping. >>Finally sort_mem: >>Was until recently left at the default of 1000. Is now 16000. > >Sort memory is per sort not per query or per connection. So depending upon >how many concurrent connections you entertain, it could take quite a chuck >of RAM. Right I understand that. How does one calculate the size of a sort? Rows * width from an explain? >>Increasing the effective cache and sort mem didn't seem to make much of a >>difference. I'm guessing the eff cache was probably raised a bit too >>much, and shared_buffers is way to high. > >I agree. For shared buffers start with 5000 and increase in batches on >1000. Or set it to a high value and check with ipcs for maximum shared >memory usage. If share memory usage peaks at 100MB, you don't need more >than say 120MB of buffers. My results from ipcs seems confusing... says its using the full 2G of shared cache: # ipcs ------ Shared Memory Segments -------- key shmid owner perms bytes nattch status 0x0052e2c1 6389760 postgres 600 2088370176 4 ------ Semaphore Arrays -------- key semid owner perms nsems status 0x0052e2c1 424378368 postgres 600 17 0x0052e2c2 424411137 postgres 600 17 0x0052e2c3 424443906 postgres 600 17 0x0052e2c4 424476675 postgres 600 17 0x0052e2c5 424509444 postgres 600 17 0x0052e2c6 424542213 postgres 600 17 0x0052e2c7 424574982 postgres 600 17 0x0052e2c8 424607751 postgres 600 17 0x0052e2c9 424640520 postgres 600 17 0x0052e2ca 424673289 postgres 600 17 0x0052e2cb 424706058 postgres 600 17 0x0052e2cc 424738827 postgres 600 17 0x0052e2cd 424771596 postgres 600 17 0x0052e2ce 424804365 postgres 600 17 0x0052e2cf 424837134 postgres 600 17 0x0052e2d0 424869903 postgres 600 17 0x0052e2d1 424902672 postgres 600 17 0x00018d45 505544721 root 777 1 ------ Message Queues -------- key msqid owner perms used-bytes messages >>What can I do to help determine what the proper settings should be and/or >>look at other possible choke points. What should I look for in iostat, >>mpstat, or vmstat as red flags that cpu, memory, or i/o bound? > >Yes. vmstat is usually a lot of help to locate the bottelneck. What would I be looking for here? # vmstat 2 10 procs memory swap io system cpu r b w swpd free buff cache si so bi bo in cs us sy id 0 0 0 1092 14780 198120 1467164 0 0 0 0 0 0 0 0 0 0 0 0 1092 19488 198120 1467204 0 0 0 0 240 564 11 5 84 0 0 0 1092 19520 198120 1467300 0 0 0 210 443 1094 29 8 63 0 0 0 1092 15832 198120 1467356 0 0 4 110 368 1455 27 5 68 3 0 0 1092 10956 198120 1467464 0 0 4 336 417 1679 33 10 57 1 0 0 1092 17840 198124 1465980 0 0 200 334 581 1914 63 14 23 1 0 0 1092 16556 198124 1466012 0 0 0 226 397 1069 30 4 66 0 0 0 1092 19096 198124 1466028 0 0 0 160 230 314 12 2 86 2 0 1 1092 16100 198128 1466748 0 0 28 1484 711 1578 23 12 65 0 0 0 1092 20140 198128 1466780 0 0 0 414 291 746 15 8 77 I'm guessing what I should look at is the io: bi & bo ? when I run some particularly large queries I see bo activity so I'm speculating that that means its reading pages from disk, correct? >>DB maintenance wise, I don't believe they were running vacuum full until >>I told them a few months ago that regular vacuum analyze no longer cleans >>out dead tuples. Now normal vac is run daily, vac full weekly >>(supposedly). How can I tell from the output of vacuum if the vac fulls >>aren't being done, or not done often enough? Or from the system tables, >>what can I read? > >In 7.4 you can do vacuum full verbose and it will tell you the stats at >the end. For 7.3.x, its not there. > >I suggest you vacuum full database once.(For large database, dumping >restoring might work faster. Dump/restore and vacuum full both lock the >database exclusively i.e. downtime. So I guess faster the better for you. >But there is no tool/guideline to determine which way to go.) Ok they had not done a full vacuum in a long time. I them run vacuumdb --full --analyze --verbose and dump it into a file. What should I look for to see if it was useful? for example: INFO: Pages 118200: Changed 74, reaped 117525, Empty 0, New 0; Tup 575298: Vac 11006, Keep/VTL 0/0, UnUsed 2454159, MinLen 68, MaxLen 1911; Re-using: Free/Avai l. Space 774122944/774122944; EndEmpty/Avail. Pages 0/118200. CPU 9.41s/1.33u sec elapsed 97.35 sec. Is there any documentation on what those numbers represent? Also do we need to use REINDEX on the indexes, or does vacuum full take case of that? >>Is there anywhere else I can look for possible clues? I have access to >>the DB super-user, but not the system root/user. > >Other than hardware tuning, find out slow/frequent queries. Use explain >analyze to determine why they are so slow. Forgetting to typecast a where >clause and using sequential scan could cost you lot more than mistuned >postgresql configuration. Right. One example I can think of is one particular query takes about 120 seconds to run (explain analyze), but if I set enable_seqscan to off, it takes about 10 seconds. >>Thank you for your time. Please let me know any help or suggestions you >>may have. Unfortunately upgrading postgres, OS, kernel, or re-writing >>schema is most likely not an option. > >I hope you can change your queries. For the most part we're not having too much trouble, just some newer queries were building for some new features is what we're seeing trouble with. >HTH > > Shridhar
On Thu, 2004-05-13 at 14:42, Doug Y wrote: > We don't seem to be swapping much: > Linux aggressively swaps. If you have any process in memory which is sleeping a lot, Linux may actively attempt to page it out. This is true even when you are not low on memory. Just because you see some swap space being used, does not mean that your actively running processes are causing your system to swap. I didn't catch what kernel version you are running, so I'm tossing this out there. Depending on the kernel (I believe 2.6+, but there may be something like it in older kernels) that you are running, you can attempt to tune this buy setting a value of 0-100 in /proc/sys/vm/swappiness. The higher the number, the more aggressive the kernel will attempt to swap. Some misc. kernel patches attempt to dynamically tune this parameter. For a dedicated DB server, a higher number will probably be better. This is because it should result in the most cache being available to the system. This, of course means, you may have to wait an tad bit long when you ssh into the system, assuming sshd was swapped out. I think you get the idea. > Swap: 1052248K av, 1092K used, 1051156K free 1465112K cached > > looks like at some point it did swap a little, but from running vmstat, I > can't seem to catch it actively swapping. > Chances are, you have some dormant process which is partially or completely paged out. For an interesting read on Linux and swapping, you can find out more here: http://kerneltrap.org/node/view/3080. Cheers! -- Greg Copeland, Owner greg@copelandconsulting.net Copeland Computer Consulting 940.206.8004