On Jul 7, 2008, at 10:24 AM, Phil Carns wrote:
There are a couple of options for dealing with the case where
num_groups > num_dfiles if we want to keep the current defaults:
1) Transparently reduce the num_groups within the distribution
calculation. From the user's perspective the params wouldn't
change; the distribution would just automatically fall back to
acting like simple stripe. We would just need to add this to the
top of a few of the dist functions:
if(num_groups > server_ct)
{
num_groups = server_ct;
}
2) Return an explicit error at create time. The varstrip
distribution has this behavior now; it actually prevents you from
creating a file if the dfile count isn't plausible for the
distribution parameters. The way to do this is to implement the
get_num_dfiles() function in twod_stripe so that it can perform a
safety check and return an error at create time if the user tries to
use too high of a num_groups setting. On the command line you would
get a generic errno (EINVAL or whatever) but we could put a
descriptive message in the logs.
Option 1) seems reasonable to me; it is a small code change, we get
to keep the existing desirable default parameters, and the single
server case just transparently falls back to simple stripe.
There's a third option, and what I've been proposing -- instead of
changing the actual parameter, adjust the value during the
logical_to_physical calculation.
-sam
-Phil
Rob Ross wrote:
I'm glad we agree that we should make the distribution work with
default parameters, and I agree there is more than one way to
accomplish that.
If you can find a way to have it default to something more
sensible, great. If not, I'm happy to have it behave like simple
stripe if the user doesn't do any tuning.
-- Rob
On Jul 4, 2008, at 2:21 PM, Sam Lang <[EMAIL PROTECTED]> wrote:
Having it work with default parameters doesn't necessarily mean
making those default parameters exactly like simple stripe. Our
best all around solution is to fix this bug with a simple check in
the 2d code and set the default parameters to something reasonable
for a 2d distribution in the general case. We looked at this last
year and decided that num_groups=2 and striping factor = 256 were
reasonable, maybe they're not, but it doesn't make sense to make
them identical to the simple stripe dist.
-sam
On Jul 1, 2008, at 8:08 PM, Rob Ross wrote:
I agree that it's best to have it work for everyone with default
parameters.
-- Rob
On Jul 1, 2008, at 4:15 PM, "Kyle Schochenmaier" <[EMAIL PROTECTED]
> wrote:
We should use simple stripe in that case, which is why i have it
defaulted to it now, someone was playing with twod-dist today and
noticed that it defaults to having 2 groups, which causes errors
when
you run it on only a single server. I would much rather have the
defaults allow the distribution to work on *any* setup, rather
than
just a n>1 servers setup.
Its not particularly useful, but I belive it is the right thing
to do,
maybe I'm wrong?
(the original default settings are worthless for my applications
so I
have to change them anyhow)
Kyle
On Tue, Jul 1, 2008 at 4:04 PM, Sam Lang <[EMAIL PROTECTED]>
wrote:
Kyle,
How is this patch useful for you guys? With a single server
(and only one
stripe), it doesn't matter what the factor and num_groups are,
does it? It
still behaves the same as simple-stripe in that case. With
multiple servers
setting the factor and num_groups to 1 does make the 2d
distribution act
like simple-stripe, but why not just use simple stripe in that
case?
-sam
On Jul 1, 2008, at 3:19 PM, Kyle Schochenmaier wrote:
Patch is against cvs head, and modifies the default num_groups
to be 1
and the group_strip_factor = 1 so twod-stripe's defaults
still work
when you use it on a single-server, it will now by default
behave the
same as simple-stripe until the user modifies the num_groups and
strip_factor xattrs.
Also added a brief description of the definitions for the
dist_params
in twod-stripe.
Tested fully functional on my servers.
Kyle
Index: src/io/description/dist-twod-stripe.c
=
=
=
================================================================
RCS file: /anoncvs/pvfs2/src/io/description/dist-twod-stripe.c,v
retrieving revision 1.5
diff -r1.5 dist-twod-stripe.c
6a7,15
/* twod-stripe will take all of the servers in the filesystem
and
* partition them into num_groups groups. Data will then be
striped to
* each group before we move onto the next group. The
strip_factor will
* determine how many chunks of strip_size are written to each
server
* in each group before we transition to the next group.
* The striping on the group level is done round-robin in the
same
* fashion as simple-stripe
*/
Index: include/pvfs2-dist-twod-stripe.h
=
=
=
================================================================
RCS file: /anoncvs/pvfs2/include/pvfs2-dist-twod-stripe.h,v
retrieving revision 1.4
diff -r1.4 pvfs2-dist-twod-stripe.h
17c17
< #define PVFS_DIST_TWOD_STRIPE_DEFAULT_GROUPS 2
---
#define PVFS_DIST_TWOD_STRIPE_DEFAULT_GROUPS 1
19c19
< #define PVFS_DIST_TWOD_STRIPE_DEFAULT_FACTOR 256
---
#define PVFS_DIST_TWOD_STRIPE_DEFAULT_FACTOR 1
-
Kyle Schochenmaier
<twod-dist.patch>_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
--
Kyle Schochenmaier
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
