Thread: Weird disk write load caused by PostgreSQL?
I have a production PostgreSQL instance (8.1 on Linux 2.6.15) that seems to be writing data to disk at rates that I think are disproportional to the update load imposed on the database. I am looking for ways to determine the cause of this I/O. As an example, here is a typical graph produced by Munin: http://purefiction.net/paste/pg-iostat.png Running an hourly iostat produces this output: Device: tps MB_read/s MB_wrtn/s MB_read MB_wrtn sda 43.50 0.21 0.04 427915 72736 sda 43.62 0.21 0.04 428183 88904 sda 43.74 0.21 0.05 428440 104877 sda 43.90 0.21 0.06 428808 124681 sda 44.06 0.21 0.07 429111 145447 sda 44.27 0.21 0.08 429532 170317 sda 44.46 0.21 0.09 429985 193594 In other words, it's reading about 400MB/hour and writing around 15-20GB/hour, or exactly 118GB during the last six hours. To determine how well this correlates to the actual inserts and updates being performed on the database, I ran a loop alongside iostat that executed "select sum(n_tup_upd), sum(n_tup_ins) from pg_stat_all_tables" against PostgreSQL every hour and output the difference. Here are a few samples: | delta_upd | delta_ins | +-----------+-----------+ | 7111 | 2343 | | 7956 | 2302 | | 7876 | 2181 | | 9269 | 2477 | | 8553 | 2205 | For the write numbers to match the tuple numbers, each updated/ inserted tuple would have to average at least 1.5MB (15 GB divided by 10,000 tuples), which is not the case; the total size of the raw tuples updated/inserted during the above session probably does not exceed a couple of megabytes. Even considering overhead, page size, MVCC, etc., this does not compute. I have not narrowed this explicitly down to PostgreSQL, but since the stock Linux kernel we are running does not offer per-process I/O statistics, I cannot determine this for sure. However, except for the PostgreSQL database, everything else on the box should be identical to what we are running on other boxes, which are not exhibiting the same kind of load. Oh, and I have PostgreSQL logging turned off. Note that PostgreSQL's performance in itself seems fine, and according to top/ps it's only very rarely in iowait. Alexander.
Alexander Staubo <alex@purefiction.net> writes: > I have a production PostgreSQL instance (8.1 on Linux 2.6.15) that > seems to be writing data to disk at rates that I think are > disproportional to the update load imposed on the database. I am > looking for ways to determine the cause of this I/O. Are you sure that iostat is to be trusted? The read numbers in particular look suspiciously uniform ... it would be a strange query load that would create a read demand changing less than 1% from hour to hour, unless perhaps that represented the disk's saturation point, which is not the case if you're not seeing obvious performance problems. regards, tom lane
On Oct 2, 2006, at 17:50 , Tom Lane wrote: > Alexander Staubo <alex@purefiction.net> writes: >> I have a production PostgreSQL instance (8.1 on Linux 2.6.15) that >> seems to be writing data to disk at rates that I think are >> disproportional to the update load imposed on the database. I am >> looking for ways to determine the cause of this I/O. > > Are you sure that iostat is to be trusted? No. :) But iostat reads directly from /dev/diskstats, which should be reliable. Of course, it still doesn't say anything about which process is doing the writing; for that I would need to install the atop kernel patches or similar. ... > The read numbers in > particular look suspiciously uniform ... it would be a strange > query load that would create a read demand changing less than 1% > from hour to hour, unless perhaps that represented the disk's > saturation point, which is not the case if you're not seeing > obvious performance problems. They are not uniform at all -- they correlate perfectly with the web traffic; it just so happens that the samples I quoted were from peak hours. Take a look at the Munin graph. (The spikes correspond to scheduled maintenance tasks like backups.) Alexander.
I'm surprised that I have to reply to myself, since in hindsight this should be bloody obvious: It's the pgsql_tmp directory. I just monitored the file creation in that directory, and found PostgreSQL to be creating huge temporary, extremely short-lived files ranging from 1MB to 20MB in size. I increased work_mem to 25MB -- should be perfectly safe on a box with 4GB of RAM, I hope - and the iostat traffic seems to have dropped to near zero. I'm a bit baffled as to why PostgreSQL would ever be sorting 20MB of data in the first place, even with ~12 connections running queries concurrently, but I suppose I will have to look more closely at our query patterns. Alexander. On Oct 2, 2006, at 19:25 , Alexander Staubo wrote: > On Oct 2, 2006, at 17:50 , Tom Lane wrote: > >> Alexander Staubo <alex@purefiction.net> writes: >>> I have a production PostgreSQL instance (8.1 on Linux 2.6.15) that >>> seems to be writing data to disk at rates that I think are >>> disproportional to the update load imposed on the database. I am >>> looking for ways to determine the cause of this I/O. >> >> Are you sure that iostat is to be trusted? > > No. :) But iostat reads directly from /dev/diskstats, which should > be reliable. Of course, it still doesn't say anything about which > process is doing the writing; for that I would need to install the > atop kernel patches or similar. > > ... >> The read numbers in >> particular look suspiciously uniform ... it would be a strange >> query load that would create a read demand changing less than 1% >> from hour to hour, unless perhaps that represented the disk's >> saturation point, which is not the case if you're not seeing >> obvious performance problems. > > They are not uniform at all -- they correlate perfectly with the > web traffic; it just so happens that the samples I quoted were from > peak hours. Take a look at the Munin graph. (The spikes correspond > to scheduled maintenance tasks like backups.) > > Alexander. > > > ---------------------------(end of > broadcast)--------------------------- > TIP 9: In versions below 8.0, the planner will ignore your desire to > choose an index scan if your joining column's datatypes do not > match