On Thu, Jul 31, 2008 at 01:33:50PM +0100, Jamie Lokier (jamie@shareable.org) wrote:
quoted text
> This is why marking the pages COW would be better.  Automatic!
> There's no need for a notification, merely letting go of the page
> references - yes, the hardware / TCP acks already do that, no locking
> or anything!  :-)  The last reference is nothing special, it just means
> the next file write/truncate sees the count is 1 and doesn't need to
> COW the page.

It depends... COW can DoS the system: consider attacker who sends a
page, writes there, sends again and so on in lots of threads. Depending
on link capacity eventually COW will eat the whole RAM.

quoted text
> > There was a linux aio_sendfile() too. Google still knows about its
> > numbers, graphs and so on... :)
> 
> I vaguely remember it's performance didn't seem that good.

<q>
Benchmark of the 100 1MB files transfer (files are in VFS already) using
sync sendfile() against aio_sendfile_path() shows about 10MB/sec
performance win (78 MB/s vs 66-72 MB/s over 1 Gb network, sendfile
sending server is one-way AMD Athlong 64 3500+) for aio_sendfile_path().
</q>

So, it was really better that sync sendfile :)

quoted text
> One of the problems is you don't really want AIO all the time, just
> when a process would block because the data isn't in cache.  You
> really don't want to be sending *all* ops to worker threads, even
> kernel threads.  And you preferably don't want the AIO interface
> overhead for ops satisfied from cache.

That's how all AIO should work of course. We are getting into a bit of
offtopic, but aio_sendfile() worked that way as long as syslets,
although the former did allocate some structures before trying to send
the data.

quoted text
> Syslets got some of the way there, and maybe that's why they were
> faster than AIO for some things.  There are user-space hacks which are
> a bit like syslets.  (Bind two processes to the same CPU, process 1
> wakes process 2 just before 1 does a syscall, and puts 2 back to sleep
> if 2 didn't wake and do an atomic op to prove it's awake).  I haven't
> tested their performance, it could suck.

Looks scary :)
Thread allocation in userspace is rather costly operations compared to
syslet threads in kernelspace. But depending on IO pattern this may or
may not be a noticeble factor... It requires testing and numbers.

-- 
	Evgeniy Polyakov
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/