I've noticed that opening a file in linux and setting the file descriptor to non blocking does not appear to work the way I would expect, I am seeing read and write calls block for up to 6 seconds waiting for a 65K write to complete. The aggregate IO rate for the RAID 0 I am doing IO to is reasonable, and the measurement mentioned above is rare, but I am using non blocking IO because I am in my server's execution state machine so I cannot afford to block that long. I have noticed that sockets under Linux work the way I expect. My solution will be to defer IO to separate threads so that I can maintain responsive service to clients. I was wondering if other developers have experienced similar issues with file IO under Linux and what solutions they may have come up with.

Brian

Are you talking about POSIX asynchronous I/O (aio_read(), aio_write(), etc)?

I think this may help what you are trying to do.

Is that the problem you're having?

Yes, I'm not sure where the quote came from, but setting the O_NONBLOCKING flag has no effect on the descriptor of a regular file. What still bothers me is that I am seeing the write function block for up to 30 seconds for an 8000 byte write! Last night I saw roughly 15 times where a small write took over 10 seconds up to 30 seconds on one of my servers. This measurement is wrapped around a write function call. I can't help but think there is something very wrong here. Unfortunately, my background is in device drivers, so I'm used to being on the other end of the call.

The quote came from man 2 open. I've never seen a write() function block noticeably on an 8k write. Are mutexes maybe involved somehow?

Thanks for the reply, I've been a way for a while. Last night I had a 40K (thats K not MB!) write take 38 seconds. The measurements I am getting are wrapped around a write call, so synchronization (at least at the application level) should not be the issue. The partition is configured as a RAID 0, and we are writing to 2-4 SATA drives. We are running Fedora core 2. The aggregate IO rate averages out to a reasonable value (30-60 MBS), but occasionally I see this anomaly. I'm concerned that using aio_write, or deferring my disk IO to separate threads may just be band-aiding a system level flaw in our overall architecture.