[Linux-cluster] GFS and server performance = Application

[isplist at logicore.net]

Hi Gordan,

The point of the test was just to get a starting point since I didn't have 
one. Even a rudimentary starting point is better than none. 
I found the 'requests per second' interesting between the applications and not 
wether they were being served up from a GFS partition or not. I have no doubt 
that GFS plays a role in performance loss but it certainly would not be as 
great a difference as I am seeing between applications. So it seems, for now 
at least.

>Is this a single-threaded or a multi-threaded test,

 ab -k -n 100 -c 100 http://192.168.1.150/ (pointing to LVS server)

>and how is the node access distribution handled?

You'll have to ask me this one in English since this is just a part time thing 
for me, I'm not interested in becoming a guru at GFS/Cluster suites. I just 
need to understand it enough to make it work for my needs. Can you rephrase 
this please?

> GFS will primarily add latency (because locks
> need to be moved between the nodes). Once the node that needs to answer
> obtains the locks, it should be able to deliver full speed on data
> transfers. If you are accessing lots of small files, the latency will be
> very dominant to the bandwidth. This could be what you are seeing.

I didn't shut the GFS or cluster services down, I only unmounted the shared 
storage for the testing. Also, there was another GFS still mounted to that 
same machine but it was in another path so not part of the path to test the 
web server. 
 
> You also don't appear to have posted the results for the same app running
> off the local FS.

Do you mean do the same test on the web server itself? 
I didn't run the test from the web server itself because I wanted a more 
accurate load. Running the tool on the server would add it's own resource 
needs to the mix. Do you think I should run it on the same server?

Just in case you do, here are those results;

This is with GFS mounted as web root and Joomla application as test page;
# ab -k -n 100 -c 100 http://localhost/

Time taken for tests:   57.43335 seconds
Requests per second:    1.75 [#/sec] (mean)
Time per request:       57043.335 [ms] (mean)
Time per request:       570.433 [ms] (mean, across all concurrent requests)
Transfer rate:          26.03 [Kbytes/sec] received

This is without GFS mounted as web root and simple index.html test page;
# ab -k -n 100 -c 100 http://localhost/

Time taken for tests:   0.187115 seconds
Requests per second:    534.43 [#/sec] (mean)
Time per request:       187.115 [ms] (mean)
Time per request:       1.871 [ms] (mean, across all concurrent requests)
Transfer rate:          267.22 [Kbytes/sec] received

> It could be down to how many files it accesses and how often. I've never
> used it, so I don't know. Remember that lock migration is expensive and
> requires cache flushes which will largely negate the usefulness of
> caching.

Same test without GFS mounted and with a basic LAMP application in root of web 
server;
# ab -k -n 100 -c 100 http://localhost/

Time taken for tests:   0.241493 seconds
Requests per second:    414.09 [#/sec] (mean)
Time per request:       241.493 [ms] (mean)
Time per request:       2.415 [ms] (mean, across all concurrent requests)
Transfer rate:          215.33 [Kbytes/sec] received

Same test without GFS mounted and with Joomla pplication in root of web 
server;
# ab -k -n 100 -c 100 http://localhost/

Time taken for tests:   35.653875 seconds
Requests per second:    2.80 [#/sec] (mean)
Time per request:       35653.873 [ms] (mean)
Time per request:       356.539 [ms] (mean, across all concurrent requests)
Transfer rate:          26.11 [Kbytes/sec] received

Of course, this test still might not mean much because it is possible that 
Joomla handles things in some unusual way. I'll have to post the same in the 
Joomla forums and see what I get back. However, right now at least, it looks 
like it is the application which makes a bigger difference than much else, 
right off the top at least.

I think I'm forgetting more version of this test? Remember that the machine 
being tested does have a second GFS mount which is always mounted during this 
testing.

If you have more thoughts on what I can test, I am more than willing to try 
them out.

Mike