Re: [fuse-devel] multi-threaded Fuse and Writes

[Miklos S. <mi...@sz...>]

Michael McTernan <Mic...@cs...> writes:

> On 27/05/11 21:53, Nikolaus Rath wrote:
>> John Haechten <jha...@pu...> writes:
>>> Is there an option for FUSE to enforce Sequential WRITE ?
>> 
>> If you mean per file: browse the ml archive, I think this was discussed
>> a few days/weeks ago.
>
> Yes - I previously submitted a patch which could provide a filesystem
> threading option where by operations taking the struct fuse_file_info
> parameter are serialised but everything else is threaded as before.
>
> My motivation for that patch was exactly for filesytems like mine and [I
> guess] John's where threads accelerates parallel access to different
> resources, but non-sequential reads or writes to a single resource have
> a penalty.  I also checked mp3fs and found it currently opts for single
> threaded access but could benefit from the threading model I patched on
> systems with more than 1 CPU (which is most computers these days).
>
> Also when using the highlevel API, as soon as you opt for threaded
> operations, you can get non-sequential access even from a purely serial
> reader or writer and this can't be fixed without going to the lowlevel
> API.
>
> Working with the highlevel FUSE API is certainly easier and while my
> patch may not have been perfect, I think there is a little bit more
> evidence that a the threading option I proposed could be beneficial to
> FUSE filesystem writers.

Yeah, I'll look into that patch once more.

Thanks,
Miklos