1 Filename: 168-reduce-circwindow.txt
2 Title: Reduce default circuit window
3 Author: Roger Dingledine
13 We should reduce the starting circuit "package window" from 1000 to
14 101. The lower package window will mean that clients will only be able
15 to receive 101 cells (~50KB) on a circuit before they need to send a
16 'sendme' acknowledgement cell to request 100 more.
18 Starting with a lower package window on exit relays should save on
19 buffer sizes (and thus memory requirements for the exit relay), and
20 should save on queue sizes (and thus latency for users).
22 Lowering the package window will induce an extra round-trip for every
23 additional 50298 bytes of the circuit. This extra step is clearly a
24 slow-down for large streams, but ultimately we hope that a) clients
25 fetching smaller streams will see better response, and b) slowing
26 down the large streams in this way will produce lower e2e latencies,
27 so the round-trips won't be so bad.
31 Karsten's torperf graphs show that the median download time for a 50KB
32 file over Tor in mid 2009 is 7.7 seconds, whereas the median download
33 time for 1MB and 5MB are around 50s and 150s respectively. The 7.7
34 second figure is way too high, whereas the 50s and 150s figures are
37 The median round-trip latency appears to be around 2s, with 25% of
38 the data points taking more than 5s. That's a lot of variance.
40 We designed Tor originally with the original goal of maximizing
41 throughput. We figured that would also optimize other network properties
42 like round-trip latency. Looks like we were wrong.
46 Wherever we initialize the circuit package window, initialize it to
47 101 rather than 1000. Reducing it should be safe even when interacting
48 with old Tors: the old Tors will receive the 101 cells and send back
49 a sendme ack cell. They'll still have much higher deliver windows,
50 but the rest of their deliver window will go unused.
52 You can find the patch at arma/circwindow. It seems to work.
56 Tor 0.0.0 through 0.2.1.19 have a bug where they only send the sendme
57 ack cell after 101 cells rather than the intended 100 cells.
59 Once 0.2.1.19 is obsolete we can change it back to 100 if we like. But
60 hopefully we'll have moved to some datagram protocol long before
61 0.2.1.19 becomes obsolete.
63 3.2. What about stream packaging windows?
65 Right now the stream packaging windows start at 500. The goal was to
66 set the stream window to half the circuit window, to provide a crude
67 load balancing between streams on the same circuit. Once we lower
68 the circuit packaging window, the stream packaging window basically
71 We could leave it in -- it isn't hurting much in either case. Or we
72 could take it out -- people building other Tor clients would thank us
73 for that step. Alas, people building other Tor clients are going to
74 have to be compatible with current Tor clients, so in practice there's
75 no point taking out the stream packaging windows.
77 3.3. What about variable circuit windows?
79 Once upon a time we imagined adapting the circuit package window to
80 the network conditions. That is, we would start the window small,
81 and raise it based on the latency and throughput we see.
83 In theory that crude imitation of TCP's windowing system would allow
84 us to adapt to fill the network better. In practice, I think we want
85 to stick with the small window and never raise it. The low cap reduces
86 the total throughput you can get from Tor for a given circuit. But
87 that's a feature, not a bug.
91 How do we know this change is actually smart? It seems intuitive that
92 it's helpful, and some smart systems people have agreed that it's
93 a good idea (or said another way, they were shocked at how big the
94 default package window was before).
96 To get a more concrete sense of the benefit, though, Karsten has been
97 running torperf side-by-side on exit relays with the old package window
98 vs the new one. The results are mixed currently -- it is slightly faster
99 for fetching 40KB files, and slightly slower for fetching 50KB files.
101 I think it's going to be tough to get a clear conclusion that this is
102 a good design just by comparing one exit relay running the patch. The
103 trouble is that the other hops in the circuits are still getting bogged
104 down by other clients introducing too much traffic into the network.
106 Ultimately, we'll want to put the circwindow parameter into the
107 consensus so we can test a broader range of values once enough relays
110 5. Transition and deployment
112 We should put the circwindow in the consensus (see proposal 167),
113 with an initial value of 101. Then as more exit relays upgrade,
114 clients should seamlessly get the better behavior.
116 Note that upgrading the exit relay will only affect the "download"
117 package window. An old client that's uploading lots of bytes will
118 continue to use the old package window at the client side, and we
119 can't throttle that window at the exit side without breaking protocol.
121 The real question then is what we should backport to 0.2.1. Assuming
122 this could be a big performance win, we can't afford to wait until
123 0.2.2.x comes out before starting to see the changes here. So we have
124 two options as I see them:
125 a) once clients in 0.2.2.x know how to read the value out of the
126 consensus, and it's been tested for a bit, backport that part to
128 b) if it's too complex to backport, just pick a number, like 101, and
129 backport that number.
131 Clearly choice (a) is the better one if the consensus parsing part
132 isn't very complex. Let's shoot for that, and fall back to (b) if the
133 patch turns out to be so big that we reconsider.