Re: [take13 1/3] kevent: Core files.

From: Evgeniy Polyakov
Date: Fri Aug 25 2006 - 01:46:29 EST


On Thu, Aug 24, 2006 at 09:03:22PM +0100, Christoph Hellwig (hch@xxxxxxxxxxxxx) wrote:
> One question on the implementation of kevent_user_ctl_modify/
> kevent_user_ctl_remove/kevent_user_ctl_add: What benchmarks did you
> do to add the separate 'fastpath' with the single onstack ukevent
> structure if there are three or less events? I can't believe this
> actually helps in practice for various reasons:
>
> - you add quite a lot of icache footprint by duplicating all this code
> - kmalloc is really fast
> - two or three small copy_from/to_user calls are quite a bit slower
> than one that covers the size of all of them.

kmalloc is really slow actually - it always shows somewhere on top
in profiles and brings noticeble overhead (as was shown in network tree
allocator project, although there were used bigger allocations).
I chose 3 ukevents, since they fit exactly one cache line (on my test
machine). In general I try to avoid allocation as much as possible, and
more generic usage case (for various servers) to accept one client and
add it, instead of waiting for several of them and commit them at once.

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