Re: How to get Web Polygraph working under Linux with 1500 IP address es

From: Alex Rousskov (rousskov@ircache.net)
Date: Sun Dec 05 1999 - 20:04:35 MST


Hi Tash,

        First of all, thanks for your useful instructions for Linux
users! I just want to address some of the questions you raise.

On Sat, 4 Dec 1999, Tash Hepting wrote:

> My test PCs are Celeron 466Mhz w/ 64MB RAM. I am using the Intel
> ProManage 100+ Ethernet card. Memory seems fine at 64MB, but the processor
> is pegged most of the time on both boxes.

"Required" CPU power with Polygraph is hard to measure. Polygraph may
use more CPU cycles than required for a given workload. In other words,
Poly's equivalent of an "idle thread" is not necessarily idle as far CPU
usage is concerned. For example, when we run 500 req/sec and 700 req/sec
experiments on our machines, the CPU utilization is very similar (close
to 100%) while both experiments run OK.
 
> run "gmake all" to build the utility. polyclt and polysrv as in the
> polygraph/src directory.

Just "gmake" is sufficient to get the essential tools (they must
compile). "Gmake all" will attempt to compile all tools and some of them
may not be supported on Linux (e.g., piper).

> Add IP aliases
> read /etc/sysconfig/network-scripts/ifup-aliases to learn about
> limitations of the aliases scripts. Remember, keep each range file within a
> single 24bit subnet mask range
> create /etc/sysconfig/network-scripts/ifcfg-eth0-range0 text file
> with to setup aliases. I have used 6 ranges, with 250 IP addresses in each
> range successfully for a total of 1500 client IP addresses.
> run "./ifup-aliases eth0" from the /etc/sysconfig/network-scripts
> directory

I am not sure, but it seem to me you are creating "permanent" aliases
here (i.e., the aliases will be preserved across restarts). If you need
temporary aliases or are changing aliases a lot, you may want to use
"aka" tool instead. For example,
        root> src/aka eth0 10.1.1-3.1-250

I am not 100% sure if "aka" works on Linux; I thought it used to.

> Add Robots
> edit the workload file and replace the IP address for the robots
> with a comma delimited list of ranges:
> '1.1.1.1-250','1.1.2.1-250','1.1.3.1-250' etc...

The above can be shortened to '1.1.1-3.1-250'.

> Be sure to increase the number of IPs for the servers also. 1500
> robots banging on one one polysrv thread doesn't work...

Hmm.. This smells like a Linux limitation or a bug in Polygraph.
Polygraph server side should run equally well with any number of IPs. If
you have a test case where fewer IPs on the server side do not work,
please report it to polygraph-bugs.

> Run the workloads
> I've got 1500 robots (1pc) requesting from 1500 servers (1pc)
> working ok in the lab with the simple.pg script and a smaller 1Byte content
> size, but occaisionally get "dst cache overflow" error on the server
> console. Not sure what this means. Smaller numbers of servers (245) don't
> generate any errors at all.

Polygraph does not emit "dst cache overflow" messages. The messages you
are seeing must be coming from the kernel or libraries. I do not know
what the message means either. Anything interesting in kernel/system
logs?

Thanks,

Alex.



This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:10 MDT