how to work on the tocs paper
Eddie Kohler
eddietwo at cag.lcs.mit.edu
Thu Mar 2 21:48:00 EST 2000
Hi all, especially Robert & Frans. This mail is about how I believe we
should work on the TOCS paper and what needs to be done.
First, where is the paper? It is in the following CVS repository:
cvs -d /home/butt/click/cvsroot co -d tocs-paper doc/tocs00
Second, what needs to be done? We need a testbed. This testbed can be a 700
MHz Pentium III or a 450 MHz Pentium III, I don't care. 450 MHz might be
easier because we could reuse some numbers from the SOSP paper. On the
other hand, we suffered i-cache misses on the 450 MHz, and did not on the
700 MHz.
Then, we need to compile and test the following Click distribution. It
consists of click-0.6 (Click as of the SOSP paper), plus the newest
ARPQuerier, plus the newest set of tools, plus the option to not use
PACKET_UPSTREAM.
cvs -d /home/butt/click/cvsroot co -d click4tocs release/zero2
I have tested this release LIGHTLY. You might need to test it heavily.
PACKET_UPSTREAM is off by default. To use PACKET_UPSTREAM, change
`config.h' to add
#define HAVE_PACKET_UPSTREAM 1
We should get the following numbers. I believe that if it is relatively
easy, we should take them all. The starred ones MUST be taken
ENTIRE CURVES, FORWARDING RATE vs. INPUT RATE
- Linux-2.2.x alone
* Linux-2.2.x + Click FromDevice w/ PACKET_UPSTREAM
* Linux-2.2.x + Click FromDevice w/o PACKET_UPSTREAM
(For the remainder of the tests, choose either w/
PACKET_UPSTREAM or w/o PACKET_UPSTREAM, whichever was
faster.)
- Linux-2.2.x + Click FromDevice SIMPLE
- Linux-2.2.x + Click FromDevice ETHERNET SWITCH
- Linux-2.2.x + Click FromDevice IPROUTER + RED
- Linux-2.2.x + Click FromDevice IPROUTER + SFQ
- Packet size vs. maximum forwarding rate, Linux & Click
* Element-by-element breakdown of Click FromDevice
(don't need to count FromDevice and ToDevice)
* Interrupt + IP processing + Device send or some similar table
* i-cache and d-cache misses
* IP processing time: Click IP router + click-xform
* IP processing time: Click IP router + click-xform + 8 nulls
* IP processing time: Click IP router + click-specialize
(to get a handle on virtual functions)
Perhaps there are other experiments we should take to get a handle on
Click's overhead.
I will help with whatever remains when I get back. I just wanted to let
this into the air, la la la
love,
ed
More information about the click
mailing list