I can't wait for direct I/O (now behind debug_io_direct setting).
Since it was NFS, you could just use tcpdump and watch what DB2 was doing on the wire. It was happily poking away sending and receiving packets all 1K in size (the current configured DB block size) with peak read and and write speeds of about 11MB/s. Since the DBAs didn't want to change settings on a production DB, I set up a testing environment, begged them to play with the direct io and block size settings on this new instance and figure out the best performance. When I checked back days later, it was set up exactly the same, "we follow best practices, use 1K block size and force direct io".
I ended up creating a VM under the guise of "we need a data warehouse" with 1/4 the cpus and ram as the DB2 machines and installed postgresql 9.2. Did a minimum amount of tuning, mostly just turning off fsync for WAL writes, then spent a week filling it up with 5TB of data and 15 billion rows from the production DB. Ran one of our analytic queries that had grown to taking 30 hours on DB2, it ran in 6 hours. The packet sizes over NFS were 32-64MB in size and getting peak speeds of 180-220MB/s on the wire.
That is not something I would suggest to people on production systems, as that would give you a good chance of data loss when the system halts. So, out of interest, were there any circumstances why turning off WAL fsync was considered a good choice in your situation?
From the docs: "setting this parameter to off does not create any risk of database inconsistency: an operating system or database crash might result in some recent allegedly-committed transactions being lost, but the database state will be just the same as if those transactions had been aborted cleanly. So, turning synchronous_commit off can be a useful alternative when performance is more important than exact certainty"
That sounds like an anecdote from the time before SSDs.
> "setting this parameter to off does not create any risk of database inconsistency: an operating system or database crash might result in some recent allegedly-committed transactions being lost, but the database state will be just the same as if those transactions had been aborted cleanly. "
If your application is told that the transactions were committed, but your DB actually hasn't, you got a problem, methinks.
Thing is DB work is all about tradeoffs, so you'll never be free of them.
As the parent explained, PostgreSQL actually lets you make a very good trade: you get full cached level performance but lose only a couple of transactions at most, instead of data corruption. If that's not ok for you, the leash can be tightened to lose nothing, but then your hardware had better keep up with your platform.
Yes, I stated "regular hard drives" :)
Note that the documentation on the config flag explicitly warns about not using it in production:
> Currently this feature reduces performance, and is intended for developer testing only.
Also note that very few things will actually do IO during commit - the only IO that I can think of are 1.) the WAL-logging of the commit (often small, a few 100 bytes at most), and 2.) replying to the COMMIT command (10s of bytes at most). It is quite unlikely that this will see much performance benefit from IO_DIRECT without further infrastructure inside PostgreSQL around io_uring and other async kernel IO apis.
From Linus himself[0]
"The thing that has always disturbed me about O_DIRECT is that the whole interface is just stupid, and was probably designed by a deranged monkey on some serious mind-controlling substances"
But back to my first question, I am still curious what your workload is that you feel will benefit from direct io :)
I use io-uring with O_DIRECT at work and the performance graphs of TLB pressure are beautiful.