Shadow and tornettools don't seem to handle packet loss well. #3325
Unanswered
higher-order
asked this question in
Q&A
Replies: 1 comment
-
Note that the lack of packet loss is just packet loss on the links, there will be packet loss from congestion even when all the links have 0 packet loss. For why you're not seeing any differences, I suspect that might have to do with this being on a 1% network. You can see the Once is Never Enough paper for why we set packet loss to 0, and an explanation for why 1% networks can behave very strangely (3.1 and 6.3, respectively). |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm simulating networks generated by tornettools, and experimenting with changing the
packet_loss
parameters on the edges in the network graphatlas_v201801.shadow_v2.gml
. The graph haspacket_loss 0.0
for all edges. I'm doing a sanity check to see if setting non-zero loss rates will indeed negatively affect the performance.The issue was first noticed when after changing all
packet_loss
to 0.011, running the otherwise same 1% Tor net for a simulation time of 60 minutes produced completely identical tornettools plots (using thetornettools plot
command). Only that the simulation took 26 hours to finish, instead of just 3 hours if allpacket_loss
are 0.0.I don't have the plots of that experiment now, but do have plots for simulations on the same 1% Tor net with
packet_loss
set to:I ran the simulations for 900 seconds each. But the plots are identical again. Except that the simulations with loss are slower.
The only other discussion I found related to "loss" is this. It is 10+ years ago and seems to deal with TCP autotuning issues.
I'd like to ask if I'm configuring loss in a wrong way, or is it more of a known issue in Shadow?
Thank you!
Edit: I've also run
iperf2
server and client on two of the virtual hosts in the network with 0.011 loss on all edges, and confirmed that it showed a much lower bandwidth (about 200x lower). But the plots are the same.Edit:
I'll upload the 3 plots in just a short while.The PDFs are uploaded.tornet.plot.pages.pdf
tornet.plot.pages.pdf
tornet.plot.pages.pdf
Beta Was this translation helpful? Give feedback.
All reactions