Here's a table of performance-related work where whole-network simulation can help:

What?CategoryExpected ImpactNext stepWhen?
#4086: TokenBucketRefillInterval valuesrate limitinghigh Figure out why graphs disagree March2012
#4682: "double door" effects on rate limitingrate limitinghigh Simulate it in #5336 March2012
#4506: N23flow controlhigh Simulate it in #4486 March2012
#4585: Read/write to fast relays more than slow onesscheduling (conn)high Need to design/write a patch
#5190: Throttle flows at guardsscheduling (conn)high Get patch from Rob
#4684: Alternate transport between relaysvarioushigh Waiting for Steven to pick one
#1854: Raise min bw for Fast flagload balancingmedium Simulators need to model relays better?
#4485: Remove stream-level sendmesflow controllow Hoping we'll like n23 so it's moot Nov2012

Below are a pile of tickets that are performance-related.

Flow control:

Compare performance of TokenBucketRefillInterval params in simulated network
Research: should N23 actually help in practice?
Research: Does N23 help especially much in cases where the first hop is slow/flaky?
Simulate the N23 design to see how it performs for typical and for really slow client connections
Decide whether letting the read bucket go negative is actually helping us

Load balancing:

Investigate raising the minimum bandwidth for getting the Fast flag
Visualize self-reported vs. measured bandwidth of relays
Graph Bw Auth Failures

Scheduling (streams, circuits, connections):

Is round-robining between each TLS conn the optimal scheduling approach?

Client time-to-bootstrap:

At what bandwidth threshold is being a dir mirror unwise?
Learn client speed trends by evaluating directory request download times

Making performance simulations more accurate:

Compare variance between simulation runs (Shadow and Experimentor)

Scaling the network to make better use of the capacity we have:

Research: what's the relationship between capacity, bytes used, and torperf?

Improving the round-trip latency, or number of round-trips, for clients:

Recognize poor guard performance and switch?
Application support for optimistic data: Torsocks

Improving performance when using bridges: No results

General / uncategorized:

Hack up stunnel to test a transport that uses a vanilla SSL handshake
Implement bwauth cap for latency
Implement bwauth cap for TCP socket exhaustion
Bwauth: Use only "slow" measurements for positive feedback
bwauth takes a long time to scan the network
BwAuth is leaving ~10% of relays unmeasured
Your Guard is failing an extremely large amounts of circuits.

Last modified 6 years ago Last modified on Mar 14, 2012, 11:05:12 PM