the performance of libeio

[changchang]

I found nodejs use libeio as file system aio lib. And the latter uses threads and synchronous io to implement aio. I read this mail list http://groups.google.com/group/nodejs/browse_thread/thread/aff97d25c59f6f2d and I think that it may be a trade off between portability and performance. But I'm worry about that file io would become the performance issues.

I try to do some test for file reading in my pc. I start 100 concurrent requests to read a file which size is 1.5M 10,000 times.
And the result of throughput is that 6MB/s per request in nodejs while 280MB/s in Java.
Does that mean nodejs is not suitable to do intensive tasks for file io?
Is there any suggestion to tune the performance of file io in practice?

Sorry for rehashing this old discussion.
Thank you and good luck :)

[bnoordhuis]

And the result of throughput is that 6MB/s per request in nodejs while 280MB/s in Java.
Does that mean nodejs is not suitable to do intensive tasks for file io?
Is there any suggestion to tune the performance of file io in practice?

It's true, file I/O is far from perfect right now.

I have some ideas on how to implement asynchronous I/O on Linux using native AIO (not POSIX AIO) but that won't be a general solution: AIO on Linux only works with files opened in O_DIRECT a.k.a. skip the page cache mode. That's acceptable for write-heavy workloads but it positively sucks for read-heavy loads because every read hits the slow, slow physical disk.

AIO is an area where Windows is likely going to run circles around the Unices...

[sh1mmer]

Thanks for the feedback. Since this isn't directly actionable I'm going to close.