Dear Phil,
Thanks for the info, I am on the dev list, so I've seen the twod_dist
debate.
They seem to have gone for the "fix on the fly" approach, where I would have
preferred the setfattr to fail with a useful error message, or have the
server
reset the client settings to ones "in bounds" so you could tell from the
client
that the settings are getting overridden. I would also like to avoid
filling the
server with error messages - as happens at the moment if there is a problem
with num_groups > number_of_servers
Thanks Again,
Tony
Tony Kew
SAN Administrator
The Center for Computational Research
New York State Center of Excellence
in Bioinformatics & Life Sciences
701 Ellicott Street, Buffalo, NY 14203
CoE Office: (716) 881-8930 Fax: (716) 849-6656
CSE Office: (716) 645-3797 x2174
Cell: (716) 560-0910 Home: (716) 874-2126
"I love deadlines, I love the whooshing noise they make as they go by."
Douglas Adams
Phil Carns wrote:
Hi Tony,
There are actually a little more text on this in trunk now. I
recently went through a similar learning experience about some of the
distributions :)
Kyle added some comments here:
http://www.pvfs.org/fisheye/browse/PVFS/src/io/description/dist-twod-stripe.c?r=1.6#l7
And I filled in some draft distribution information here, which talks
a little about incast:
http://www.pvfs.org/fisheye/browse/PVFS/doc/pvfs2-tuning.tex?r=1.2#l267
As for the 2 groups problem: that is actually not really a valid
setting for a single server case. For the 2d dist to be beneficial
the num_groups needs to be no larger than the number of servers, and
ideally evenly divisible into the number of servers. I don't know if
you are on the pvfs2-developers list, but there is an ongoing
discussion thread about the issue you brought up of the default 2d
param values not working in the single server case:
http://www.beowulf-underground.org/pipermail/pvfs2-developers/2008-July/004092.html
The short story is just use more servers or else reduce the value of
num_groups for the time being if you want to experiment with it.
-Phil
Tony Kew wrote:
Dear Phil,
Its looking good now on a single node test - I will expand to a 16
node test next...
The twod_stripe distribution is still seeing the errors about 2
groups I sent in my
last email, but the file is created anyhow.
In that regard, does someone have a good explanation of the 2D stripe
& how
it works, i.e. what the "num_groups" and "group_strip_factor" do to
change the
stripe behavior? The only comment I could find was that the
distribution is
designed to combat incast, but I can't find a reference to what
"incast" is
either!
Many, Many thanks,
Tony
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
