[#LU-728] pdsh HOSTLIST expression support in test sutie

[LU-728] pdsh HOSTLIST expression support in test sutie Created: 03/Oct/11 Updated: 27/Mar/12 Resolved: 04/Jan/12
Status:	Resolved
Project:	Lustre
Component/s:	None
Affects Version/s:	Lustre 2.2.0
Fix Version/s:	Lustre 2.2.0

Type:

Bug

Priority:

Minor

Reporter:

James A Simmons

Assignee:

Minh Diep

Resolution:

Fixed

Votes:

Labels:

None

Environment:

Lustre nodes running auster suite

Rank (Obsolete):

4792

Description

With the testing of large stripe count with the lustre test suite creating a list of OSTDEV and oss host can become very large so to simply this I implement pdsh HOSTLIST for the test suite configuration file. For example instead of ost1_HOST=foo1;ost2_HOST=foo2,etc you can do a simple ost_HOST=${ost_HOST:~~"foo[1-n]"}. Also with the power of HOSTLIST combos like ost_HOST=${ost_HOST:~~"foo[1-3,4,9-20]"} are possible. Many forms of HOSTLIST have been tested in our test beds. Can be expanded later to devices as well as host.

Comments

Comment by Peter Jones [ 05/Oct/11 ]

Yu Jian

I think that this one will be of interest to you

Peter

Comment by James A Simmons [ 05/Oct/11 ]

Patch at http://review.whamcloud.com/#change,1462

Comment by Peter Jones [ 30/Nov/11 ]

reassigning to Minh

Comment by Minh Diep [ 01/Dec/11 ]

James,

please correct me if I am wrong but the patch you upload is about RCLIENTS and nothing to do with ost_HOST. I don't believe we can apply the hostlist_expand to any of the server names which are in the form ost#_HOST = hostname. If you agree, please change the description to reflect what the patch does

Thanks

Comment by James A Simmons [ 08/Dec/11 ]

You can by using get_nodes. I will include that in the next patch. The problem with the RCLIENTS and CLIENTS list is the comma list and non comma list. Its unclear in ncli for those list to have the values seperated by a comma or not. Simple example-

ncli.sh
-----------------------------------------
CLIENT1=${CLIENT1:-`hostname`}
SINGLECLIENT=$CLIENT1

Note spoon01 is local host, should be filtered out.
RCLIENTS=${RCLIENTS:-"spoon01 spoon03"}
-----------------------------------------

From my test file

echo "CLIENTS=$CLIENTS RCLIENTS=$RCLIENTS"
FAIL_LIST=($RCLIENTS)
FAIL_NUM=${#FAIL_LIST[*]}
echo "FAIL_NUM=$FAIL_NUM"
echo "CLIENT2 = $CLIENT2"

I get...

CLIENTS=spoon01,spoon03 RCLIENTS=spoon01 spoon03
FAIL_NUM=2
CLIENT2 = spoon03

Now if you have a comma in the RCLIENTS then FAIL_NUM=1. Also `hostname` is spoon1 yet it ends up in RCLIENTS. Commas are also a issue in the simple loop

for client in $CLIENTS; do
echo "$client !!!!"
done

Which prints spoon01,spoon03 !!!! This is also wrong.

The next patch will greatly simpliy this handling.

Comment by James A Simmons [ 08/Dec/11 ]

The last patch should sort the client list out with or without commas. Also the example I gave from insanity doesn't handle the comma listing of RCLIENT well. From what I can tell all host list are expected to have commas normally.

Comment by Minh Diep [ 09/Dec/11 ]

"The problem with the RCLIENTS and CLIENTS list is the comma list and non comma list. Its unclear in ncli for those list to have the values seperated by a comma or not."

We should make it work for both.

There is also an issue with hostname return with domain name (ie. host.local) and this will not be filtered out

Comment by James A Simmons [ 12/Dec/11 ]

Great, because this patch allows both in the client list. I looked at the value of $HOSTNAME to discover that it always does not have the domain name. We could always go with the short name (hostname -s) but is that what we want? Consider the case of WAN support. On last thing that I want to fix up is facet_active. That is a lot of file at a very large scale when testing large stripe count support. Their has to be a better way.

Comment by James A Simmons [ 16/Dec/11 ]

Since get_nodes needs more work I can remove it from this patch and include it in a later patch. Looking at the code the best way to handle a node list is without commas. If we need a comma list a function exist to do that. I will posta updated patch that makes the node list without commas.