Re: [fuse-devel] fuse performance

[Mohit A. <ext...@gm...>]

Which Linux kernel version did you use for your tests ? Also, what fuse
version did you use ? I'd also appreciate advice on the FUSE options that
helped you eliminate some of the context switch overhead.

Thanks,


- Mohit


On Sun, Jul 6, 2008 at 5:23 AM, Szabolcs Szakacsits <sz...@nt...>
wrote:

>
> On Sun, 6 Jul 2008 Mohit Aron wrote:
>
> > That's very disappointing.
>
> Recently we made some tests with the unoptimized ntfs-3g and FUSE large
> write support and found that block device based bulk write can be actually
> faster than ext3 on a T9300@2,5 (665 MB/sec vs 621 MB/sec) even in the
> lower range of big files. The fastest ever speed measured with ntfs-3g
> via context switches was 802 MB/sec.
>
>  block
>  size  tmpfs   ntfs-3g  ext3
>
>  512     421      16     128
>   1k     613      31     287
>   2k     775      57     393
>   4k     898      98     545
>   8k     949     174     579
>  16k     973     289     593
>  32k     964     395     603
>  64k     971     515     613
>  128k     977     665     621
>  256k     979     661     622
>  512k     979     665     618
>   1M     977     644     625
>
> The current scalability limit in the kernel is at 128 kB block size
> (32 pages per request), so the FUSE result could be still better.
>
> Testing with fusexmp* can give bogus results because the performance
> bottlenecks are often in the underlaying in-kernel file systems.
>
> There are several, documented FUSE options which can eliminate quite a
> lot of needless context switches via kernel caching and other ways when
> this can be possible (e.g. most block device based file systems).
>
>        Szaka
>
> --
> NTFS-3G: http://ntfs-3g.org
>