Fwd: How to tune Linux VFS readahead size

Fwd: How to tune Linux VFS readahead size

[Peter Xu]

Hi, everyone,

I just want to know whether there is better way to make Linux VFS read
ahead more aggressive (from my reading the source code, the maximum
read ahead of file is statically set to 2MB). E.g., to change the read
ahead window to 64MB.

What I am doing is tuning CEPHFS read performance, however, remote is
very poor (100+MB) comparing to local test (374MB). I am testing using
dd (this should also the case for most real-life scenarios), so an
aggressive read ahead policy should be really help for me to make the
IO asynchronous and make the backend disks running concurrently.

Thanks in advance!
Peter

Re: Fwd: How to tune Linux VFS readahead size

[Jan Kara]

  Hello,

On Tue 16-06-15 16:32:18, Peter Xu wrote:
> I just want to know whether there is better way to make Linux VFS read
> ahead more aggressive (from my reading the source code, the maximum
> read ahead of file is statically set to 2MB). E.g., to change the read
> ahead window to 64MB.

No, there is no way to tune automatic readahead window to more than 2 MB.
You can still ask the kernel explicitely via posix_fadvise(2) to prefetch
the data for you though.

> What I am doing is tuning CEPHFS read performance, however, remote is
> very poor (100+MB) comparing to local test (374MB). I am testing using
> dd (this should also the case for most real-life scenarios), so an
> aggressive read ahead policy should be really help for me to make the
> IO asynchronous and make the backend disks running concurrently.

So you have some larger chunks in which file is split and you need to fetch
them from different servers, do I understand it correctly?

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR