Re: [9fans] Speed of 9pfuse on Linux

[Eric Van Hensbergen <ericvh@gmail.com>]

On Feb 10, 2010, at 7:32 AM, Pavel Klinkovsky wrote:

>> 1) real plan9 to the same place
>> 2) qemu plan9 on Fedora to the same place
> As I wrote above, I made exactly the same test on exactly the same HW
> (and internet connection).
> 1. Native Plan9.
> 2. Native Fedora 10 with p9p.
> 
>> "It's slow, what's wrong" is perhaps a little vague.
> Not precisely measured (I can do it today).
> My estimation of the time spent by 'ls' command in contrib directory:
> - 9pfuse on Fedora was more than 10 times slower.
> 

That's actually probably pretty good, if you look at slide  10 & 11 on my Linux 9P Trace and Walkthrough (top of http://www.graverobber.org at the moment) you'll see the bad protocol behavior caused by v9fs at the moment -- 9pfuse probably has around the same order increase.  The problem is that Linux doesn't know we got all the attribute information with the dirread, and then goes and individually queries each file in the directory with a stat -- this happens in a serial fashion, so the high latency to sources just makes the problem worse.

A potential solution is to cache metadata, which would speed things up dramatically for static data, but might mess things up for synthetic files.

      -eric