On Mon, 6 Nov 2000, Joe Reifel wrote:
> I was looking at the workload file for reverse mode
> proxy testing -- webaxe.pg, and had some concerns about capacity
> testing. It looks like the script is supposed to work the
> "polymix-2" way, where each client and server gets a single IP
> address, as opposed to the "polymix-3" way, where each client and
> server gets a range of IP addresses, allocated via the aka
> utility.
Webaxe-1.pg is indeed based on the PolyMix-2 workload. However, both
PolyMix-2 and PolyMix-3 use the same IP allocation scheme, one address
per simulated robot or origin server (each polyclt or polysrv gets a
range of IP addresses). The IP addresses can be "created" using aka or
any other tool/script external to Polygraph.
> Do you have any comment on whether or not we can go
> ahead and use client aliases for the webaxe workload instead of
> individual IP's?
You sure can. From Polygraph or even FreeBSD point of view, there is
no difference between "individual IP" and an alias. You will create
lots of robots (each with an IP alias) and a few origin servers (each
with an IP alias).
> And how do we get around the fact that in reverse mode, only a
> single web server is being cached?
That is not necessarily a fact. :) In general, an accelerator can
accelerate many origin servers at once. Many real configurations use
multiple origin servers behind an accelerator. For example,
money.yahoo.com and travel.yahoo.com could have been two different
origin servers being accelerated by a single reverse proxy (or a group
of proxies).
A typical setup may look like this:
clients
||
L4/7-redirector === proxies
||
origin servers
Clients think that they talk directly to the servers. A redirector
"transparently" diverts port 80 traffic to the proxies that would
accelerate the origin servers.
A setup with a single virtual IP address (VIP) can also be tested with
Polygraph, but some error and warning messages will affect such a
test. More work is needed to fully support VIPs.
> Should we just set up our cache to cache a range of web server
> IP's to accomodate the test? It seems that the key limitation
> here is that we're now talking about a single web server address,
> although almost 100% of the content, under the right
> circumstances, is designed to be cacheable, so I can see how a
> single server could probably handle a significant load. I
> understood that there had to be one client to each server though,
> and that doesn't make sense for this kind of testing, IMO.
You should use as many origin servers as required to handle the load.
You are right that the miss load on origin servers will be light
compared to PolyMix, so fewer server will be needed. You may think of
those origin servers as a single "logical" server being accelerated.
> I haven't looked at this very much yet, but I was
> hoping I could get these questions answered before I started
> investing time into setting up a test.
Hope the above helps. If you plan on using Polygraph 2.5.4, please let
me know and I will send you a more up-to-date version of the WebAxe
workload that we have been using in private tests with 2.5.4.
Alex.
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:16 MDT