Conference destination: Hilton
Grosvenor – over the road?
Hotel Royal Plaza – no rooms available, just down the road perhaps. 1905 Hotel Plaza Blvd. — Actual reservation website
Doubletree Guest Suites – further down the road maybe. 2305 Hotel Plaza Blvd. $139/night according to Expedia.
Best Western. 2000 Hotel Plaza Blvd. $80/night according to Expedia. Hotel reservation site seems a bit broken. Seems very close?
Nimrod/O is still running! Well, it was until I paused it, seemed it was slowing voodoo down quite a bit. I’ll leave it running over the weekend (must remember to resume it before I leave work).
I’ve added more and more references to my thesis and sorted them out better on citeulike. The references are slowly but surely building up, I think there are now 26. Not heaps yet, just steady progress.
I also worked a bit more on paper and then incorporated my thoughts into the section headings. I’m still not happy, but I am happier than I was with the current structure of the thesis. Needs further thought and discussion I believe.
Right, I’ve set up a fairly simple simulation with Nimrod/O again. This time it has some parameters for cross traffic. No random loss. Given that we are now simulating up to 200 TCP flows, I imagine it will take a bit longer to get results.
In other news I am researching somewhere to stay for wintersim. The Hilton is rather expensive.
Comparing FreeBSD 5 and ns-2 Agent/TCP/Sack1/Agent/TCPSink/Sack1/DelAck. In a simple dumbbell scenario with very few TCP streams.
At the end of this post is the verbatim output of Nimrod/O for this scenario (well, most of it). What it boils down to is that in the situation where:
fwd_queue=20 rvs_queue=38 fwd_delay=0.160718 rvs_delay=0.196312
fwd_loss=0.202638 rvs_loss=0.313131 fwd_bw=1281154 rvs_bw=6701321
Then we get the following result:
sim$ tail result-*
==> result-freebsd.txt <==
1351732
==> result-ns2.txt <==
3040
That is the total amount of bytes through the network achieved. Obviously ns-2 struggled to make a connection while FreeBSD actually transfered a significant amount of data (over the 10 minute simulation time). Really need to do more analysis with different random seeds and so on to test this sort of result.
And do we even care? 20% loss in one direction and 31% in another. RTT of ~ 354 ms. Upstream bw of ~1.3Mb, downstream of ~6.7Mb (Well, I think the unit is b/s). I don’t believe the queue sizes are so important here.
The result is interesting enough in itself I guess. But next I should vary the simulation script and the parameters I give it. Eg. no random loss, but more cross traffic mixes. I already know random loss provides scenarios where we get quite different results.
Also, worth testing against Linux and OpenBSD.
------------- For text variables: direction="reverse" ---------------
Starting point Optimum At Point
Iters. Evals. Batches
(46,36,0.7165,0.6616,0.04521,0.1588,5.891e+06,8.179e+06) 0.673575
(49,34,0.6994,0.6417,0.05621,0.1701,5.72e+06,8.018e+06) 50 267 (277) 58 (58)
(24,40,0.1861,0.1612,0.1853,0.2707,5.41e+05,7.697e+06) 0.00225404
(20,38,0.1607,0.1963,0.2026,0.3131,1.281e+06,6.701e+06) 23 103 (106) 24 (24)
------------ Statistics: Combination 2 Block 'Simplex' -------------
Total number of searches: 2
searches aborted: 0
searches failed: 0
successful searches: 2
Search objective results:
mean: 0.337914
standard deviation: 0.474695
standard error : 0.33566
best: 0.00225404
worst: 0.673575
Evaluations performed (scheduled):
mean: 185.00 (191.50)
standard deviation: 115.97 (120.92)
standard error: 82.00 (85.50)
maximum: 267 (277)
minimum: 103 (106)
Batches performed (scheduled):
mean: 41.00 (41.00)
standard deviation: 24.04 (24.04)
standard error: 17.00 (17.00)
maximum: 58 (58)
minimum: 24 (24)
Iterations:
mean: 36.50
standard deviation: 19.09
standard error: 13.50
maximum: 50
minimum: 23
Best Optimum=0.00225404 for parameters fwd_queue=20 rvs_queue=38
fwd_delay=0.160718 rvs_delay=0.196312 fwd_loss=0.202638 rvs_loss=0.313131
fwd_bw=1281154 rvs_bw=6701321 direction="reverse"
---------------------------------------------------------------------
-----------------------------------
-----------------------------------
-------------------- Statistics: All Experiments---------------------
Total number of searches: 4
searches aborted: 0
searches failed: 0
successful searches: 4
Search objective results:
mean: 0.337914
standard deviation: 0.387587
standard error : 0.193794
best: 0.00225404
worst: 0.673575
median: 0.337914
Iterations:
mean: 36.50
standard deviation: 15.59
standard error: 7.79
maximum: 50
minimum: 23
median: 36.5
Best Optimum=0.00225404 for parameters fwd_queue=20 rvs_queue=38
fwd_delay=0.160718 rvs_delay=0.196312 fwd_loss=0.202638 rvs_loss=0.313131
fwd_bw=1281154 rvs_bw=6701321 direction="forward"
fwd_queue=20 rvs_queue=38 fwd_delay=0.160718
rvs_delay=0.196312 fwd_loss=0.202638 rvs_loss=0.313131 fwd_bw=1281154
rvs_bw=6701321 direction="reverse"
-------------------------- Combined Counts --------------------------
Evaluations performed (scheduled):
total: 740 (766)
mean: 185.00 (191.50)
Batches performed (scheduled):
total: 164 (164)
mean: 41.00 (41.00)
---------------------------------------------------------------------
Nimrod/O is currently running a whole lot of network simulations. I’m currently comparing the amount of data successfully sent between FreeBSD5 and ns-2 FullTcp/Sack. I probably shouldn’t be using FullTcp, but it will do for now.
I have no idea of how long this is going to take. I have about 9 parameters. Previously with 4 parameters it took a couple of minutes. Still using the simplex algorithm, but with a tolerance of 0.1 instead of 0.0 now.
I’ll leave Nimrod/O running overnight and see how it goes. It is running 2 ns-2 processes concurrently, but I don’t think it really phases anyone working on voodoo. Voodoo seems to be able to cope these days.
So that is probably enough practical work for today. Maybe I’ll find a little time now to get some thesis work done since I got next to nothing done this morning.
Lovely, innit?
nimrodo -l -f sim.shd
Nimrod/O version 2.6.1
Copyright 1999-2005 by Monash University.
See the file Nimrodo_License in the installation directory
for usage conditions.
Monash University makes no warranty of any kind with regard
to the results produced by Nimrod/O.
------------------------- Optimization block 1 -----------------------------
method simplex
starts 1 named "Simplex"
starting points random
tolerance 0
endstarts
endmethod
--------------------------------------------------------------------------------
Block 'Simplex' : selecting starting points
Block 'Simplex'
no. starts 1
Start 1: starting point (46, 36, 0.716456, 0.661607)
-------------------------------------------------
Block 'Simplex' : Beginning simplex optimization for start 1
./run-ns.sh: line 3: 2600 Segmentation fault ./ns $@
./run-ns.sh: line 3: 2598 Segmentation fault ./ns $@
./run-ns.sh: line 3: 2617 Segmentation fault ./ns $@
./run-ns.sh: line 3: 2618 Segmentation fault ./ns $@
./run-ns.sh: line 3: 2627 Segmentation fault ./ns $@
m> _______________________________________
m> Start 1 complete
m> optimization failed with message 'Evaluation program failed at all 5 points in the starting simplex'
The morning was a bit of a loss. Only got around to starting to do some more work on my thesis quite late in the morning, then found that Gentoo’s tetex 3.0 wouldn’t compile my .tex files. For some reason it seemed to default to pdfetex rather than etex for the latex symlink. And I’m not sure Gentoo’s etex for tetex 3.0 even works. I eventually gave up and stuck >=app-text/tetex-3.0.0 in my /etc/portage/package.mask and emerged tetex again. Problem sorted for now!
Then went for a bike ride in weather that ended up being windy and rainy. Was good though.
Finally in uni and I’m about to try and get down and dirty with Nimrod/O. Lets see how it goes…
Exploring the parameter space of simulations with Nimrod/O sounds like a bit of fun. Nimrod/O has runs many simulation runs with varying parameters and using an optimisation algorithm to choose the next set of parameters.
So I can use it to explore differences in TCP stacks if I set it up correctly. Thanks to Tony McG for pointing this out to me.
The end-to-end working group mailing list has some very good discussion on network simulation at the moment. My current favourite email is this one that gives some opinions on why network simulation is important and some weaknesses.
After working on chapter and section headings for my thesis this morning (that was going rather slowly) I’ve looked into my simulations using my SURGE models some more.
At the moment I’ve removed connection disestablishment because that breaks FullTcp and isn’t implemented in NSC yet (though I believe that would be trivial).
It looks like this is worth investigating a bit further. The preliminary graph is quite interesting. But I need to explore the parameter space a bit more before I know what is really going on. And for some reason Linux is doing a fair amount worse than BSD on this. I suspect a bug or configuration error at this point in time.
While my wireless sims use someone else’s scripts, it would seem that they were published as example wireless testing scripts and not actually used to generate results from.
So not so important. I’ll leave this for now. Might be worth revisiting, there is a reasonable difference in some of the results after all.
Back in NZ, in the Tron to be exact.
First step is to write an abstract for my thesis. This is somewhat done after today but requires further work.
My time is now 50/50 between thesis and practical work. We’ll see how that works out. On the practical work side I am back to looking at the situations I was looking at when I left: wireless sims and so forth. Should also finish off my quick simulation sim-30.
Powered by WordPress