Opened 8 years ago

Closed 2 years ago

#4487 closed task (wontfix)

Research: Does N23 help especially much in cases where the first hop is slow/flaky?

Reported by: arma Owned by:
Priority: Medium Milestone:
Component: Metrics/Analysis Version:
Severity: Keywords: performance flowcontrol
Cc: robgjansen, tschorsch@…, iang@…, malsabah@…, intrigeri Actual Points:
Parent ID: #4506 Points:
Reviewer: Sponsor:

Description

Once upon a time I thought that using Tor on a modem shouldn't be that bad -- the modem is bad, and the Tor network is bad, but the combination of them should be dominated by whichever is worse.

Tor turns out to not work that way in practice: composing a bad first link with a bad Tor network yields a way worse experience than can be explained by either component by itself.

My hypothesis is that this crummy experience is due to the same issue that N23 aims to address: the fact that our end-to-end flow control needs cells to make it through an entire round trip before more cells are sent.

We should explore N23 simulations in the case where Alice's connection to her first hop is flaky (where flaky might be some combination of low bandwidth, high latency, and non-trivial packet loss), and for various loads on the Tor network side.

Oh, and we should probably get some control runs where we do the simulations with vanilla Tor rather than Tor+N23.

Child Tickets

Attachments (1)

20121115-apollo-n23-flaky-combined.pdf (615.8 KB) - added by robgjansen 7 years ago.

Download all attachments as: .zip

Change History (12)

comment:1 Changed 8 years ago by arma

Keywords: performance, flowcontrolperformance flowcontrol

comment:2 Changed 8 years ago by robgjansen

Cc: jansen@… added

comment:3 Changed 8 years ago by Flo

Cc: tschorsch@… added

comment:4 Changed 8 years ago by karsten

Parent ID: #4506

comment:5 Changed 8 years ago by iang

Cc: iang@… added

comment:6 Changed 8 years ago by arma

I wrote up in #5333 a wishlist item for Experimentor and Shadow that I think will help us better understand the situation (once we have an n23 patch, that is).

comment:7 Changed 7 years ago by arma

Results in #4486 show "no, n23 doesn't look like a good idea with the parameters we picked, unless there's a bug in the n23 patch or something".

Rob, is now a good time to do a run where some of the clients are on slow/flaky first hops, and see if n23 looks better in this light?

I don't have a good handle on what I mean by 'slow/flaky', so "pick whatever definition you like" is my first suggestion.

Changed 7 years ago by robgjansen

comment:8 in reply to:  7 Changed 7 years ago by robgjansen

Cc: robgjansen added; jansen@… removed

Replying to arma:

Results in #4486 show "no, n23 doesn't look like a good idea with the parameters we picked, unless there's a bug in the n23 patch or something".

Rob, is now a good time to do a run where some of the clients are on slow/flaky first hops, and see if n23 looks better in this light?

I don't have a good handle on what I mean by 'slow/flaky', so "pick whatever definition you like" is my first suggestion.

I ran 100 of the 2375 web clients as "slow" web clients, configuring their access to the rest of the network with 200 millisecond latency, 25 ms jitter, and a 15% chance that each packet gets dropped. These nodes were given bandwidths of 500 KiB/s down and 200 KiB/s up.

I again ran vanilla, 'UseN23 1', and 'N3Initial 50' configs.

The results are in these graphs. They show a bigger improvement in last byte download time for the slow web clients than for the normal web clients, matching your intuition. However, overall the performance boost is still not what we had hoped.

comment:9 Changed 7 years ago by Mashael

Cc: malsabah@… added

comment:10 Changed 5 years ago by intrigeri

Cc: intrigeri added

comment:11 Changed 2 years ago by karsten

Resolution: wontfix
Status: newclosed

Closing tickets in Metrics/Analysis that have been created 5+ years ago and not seen progress recently, except for the ones that "nickm-cares" about.

Note: See TracTickets for help on using tickets.