Messages in this thread | | | From | Steve Whitehouse <> | Subject | Re: Zerocopy NBD | Date | Wed, 30 May 2001 18:15:30 +0100 (BST) |
| |
Hi,
> > On Wed, 30 May 2001, Steve Whitehouse wrote: > > [info about NBD patch deleted] > > > Cool. > > Are you seeing performance improvements with the patch ? >
Yes, but my testing is not in anyway complete yet. The only network device I have which is supported by zerocopy is loopback and there appear to be problems with deadlocks when using NBD over loopback. So what I did was to modify the NBD server (the userland one from Pavel Machek's web site) so that it didn't actually do any disk I/O. It still copied the data from the network into a buffer on write and it returns zeroed buffers on read (not that thats important as only the write patch is affected in the patch).
I could then test using dd which is a bit artificial in that it creates large requests giving probably much more data per NBD request than would be usual under a filesystem load and hence also better with the zerocopy patch. A timed dd with 100000 blocks of 1k spent 1.2 secs of system time to do the write with NBD in 2.4.5 and 0.8 secs with my patch.
Also it may well be possible to adjust the network stack's memory management to give better performance. I upped the values in tcp_[r|w]mem but I've not checked what different vaules would do to those figures.
I want to do some more testing though in case I've made an error somewhere in the method. I'd be particularly interested to hear from someone who has any results for real hardware. If I have time I'll look into whether the eepro100 or SysKonnect GigE cards could be made to support zerocopy as they are the ones I have here,
Steve.
- 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/
| |