[Gluster-users] Gluster 3.3.0 on CentOS 6 - GigabitEthernet vs InfiniBand

Jeff Darcy jdarcy at redhat.com
Thu Oct 18 14:50:59 UTC 2012

On 10/18/2012 10:30 AM, Corey Kovacs wrote:
> Jeff that came from one of you core glusterfs engineers (sid?). According to
> him the term stripe isn't used correctly and does not imply parallel
> operations.

Well, not to pull rank or undermine a colleague, but I don't think Sid knows
the code like I do.  Go look at stripe_writev if you want to see how it will
issue multiple requests within a loop xlators/cluster/stripe/src/stripe.c:3754
before waiting for any to return.  You can even use gdb to step through it if
you still don't believe me.

> Writes are handled one brick at a time. If a stripe were parallel
> tje my 800 Mbs should have been closer to 2GB across three servers each capable
> of 900 MB/s all day long.

I agree on "should" but there are other explanations besides lack of
parallelism.  Please note that diverting people to the wrong place keeps them
from fixing anything in the right place.

More information about the Gluster-users mailing list