Version 171 (modified by trac, 9 years ago) (diff)


Hacking Firefox for Maximum Performance with Tor


Tor is known for being secure but slow. If you want to improve browsing speed a bit, please follow the following simple instructions for tweaking the Firefox web browser's settings:

Table of contents

  • Procedure 1
  • Procedure 2
  • The proof is in the pudding, results
  • Advanced Tuning for Windows

Procedure 1

First, open Firefox's advanced settings menu by running about:config from the address bar. Upon entering this address, you will see a long list of internal settings. Modify the following ones and set them to the suggested values shown here for maximum performance:

network.http.keep-alive.timeout:600 (300ms default is OK usually, but 600 is better.)
network.http.max-persistent-connections-per-proxy:16 (Default is 4)
network.http.pipelining:true (Default- false. Some old HTTP/1.0 servers can't handle it.)
network.http.pipelining.maxrequests:8 (No default)
network.http.proxy.keep-alive:true (Default- true, but double check)
network.http.proxy.pipelining:true (Default- false) - See NOTE1 below.}}}

Afterwards, just restart the browser and experience the difference! For some automated additional performance hacks, check out the [[|FasterFox extension]]. You also can do the same tweaks manually with the help of [[|this page]].

NOTE1: Proxy pipelining may not be well supported by Privoxy. For this reason, you may want to install [[|Polipo]] and use that instead of Privoxy to get the performance benefits of pipelining. If you use [[|Torbutton]] (which you should, if you want [[|any anonymity at all]]), all of the Tor-relevant privacy scrubbing features of Privoxy are no longer necessary.

NOTE2: Do not use page prefetching. Disable this if it is enabled. Prefetching is a speculative feature, which assumes that you will read the pages referenced by the links in the current page you are viewing. This places undue load on the Tor network and clog your circuits with unnecessary traffic. Its unlikely you will read all the pages referenced by the current page, especially in the case of search engines results.

== Procedure 2 - A Tor Non-Functional Requirement (NFR) ==
If you follow the previous authors work you should have well performing access. To go that bit further lets consider the ideal behaviour of our Tor client.

You will need: [[|The on-line reference to Tor properties, that can be placed in torrc.]] Always back up this file before editing.

Lets think of a Non-Functional Requirement we might like to place on our Tor client.

 * we want it to establish circuits as quickly as possible. If it takes too long to do this ignore them, by timing out the building of circuits quickly.
 * now we have circuit build time-outs occuring more frequently, we need to encourage Tor to try to generate circuits more often.
 * Once we have established a circuit, we are assuming its a good one and we dont want it being timed out by firewalls or anything else. We need to make sure a ping occurs on the circuit to prevent this.
Given this NFR, lets come up with some properties that may help satisfy it.

 * CircuitBuildTimeout NUM
  . Try for at most NUM seconds when building circuits. If the circuit isn't open in that time, give up on it. (Default: 1 minute.) Force circuits that are quick to establish and thus likely to push traffic more quickly. Values as low as 2 seconds have been tried with good results, although this can cause severe damage to the Tor network if your network connection is simply not fast enough to establish any circuits in this time. The effect is a smaller 'Topological Radius' of servers used for Tor, ie the network connections available from your connection. Unfortunately, the smaller you make this number, the smaller the number of paths your client will use, and the less your anonymity.
  As of Tor v0.2.1.9-alpha, the CircuitBuildTimeout has been clipped to 30 seconds, which affects some of the tuning recommendations here. From the release notes:"
  ''Clip the CircuitBuildTimeout to a minimum of 30 seconds.
  Warn the user if lower values are given in the configuration.
  Bugfix on Patch by Sebastian.''"
 * NumEntryGuards NUM
  . If we are going to be decreasing the CircuitBuildTimeout, you want to increase the likelihood you have a guard node fast enough to build these fast circuits for you. NUM=5 to 8 are good choices here.
 * KeepalivePeriod NUM
  . To keep firewalls from expiring connections, send a padding keepalive cell every NUM seconds on open connections that are in use. If the connection has no open circuits, it will instead be closed after NUM seconds of idleness. (Default: 5 minutes)
 * NewCircuitPeriod NUM
  . Every NUM seconds consider whether to build a new circuit. (Default: 30 seconds) Lets make Tor ready to establish a new circuit more readily.

Settings that you may append to the end of the torrc configuration file are as follows:
# Try for at most NUM seconds when building circuits. If the circuit
# isn't open in that time, give up on it. (Default: 1 minute.)
CircuitBuildTimeout 5

# Increase the number of guards to increase the likelihood that 
# you will have a few guards fast enoiugh to build these circuits.
NumEntryGuards 6

# Send a padding cell every N seconds to keep firewalls from closing
# our connections while Tor is not in use. (Default: 5 minutes)
KeepalivePeriod 60

# Force Tor to consider whether to build a new circuit every NUM
# seconds. (Default: 30 seconds)
NewCircuitPeriod 15

As an added bonus, with these settings the "New Identity" button in Vidalia also effectively becomes a "This circuit is WAY too slow" button that you can mash to get a faster, more responsive circuit if Tor ever starts choking up on you. This is necessary because sometimes circuits will start out fast, but then get overloaded as traffic from other clients bursts and overwhelms the slowest node in the path.

== The proof is in the pudding, results ==
With the changes made from Procedure 1 and 2, and a 2Mb connection, you can realise a sustained throughput of >100k, peaking at ~256k or more, with a ping response time of between 250 and 900ms. In other words, experimentation away from default settings, should enable you to saturate your internet connection (depending on Tor node capacity), which is the ideal result. All results tested under Tor v0.1.2.7-alpha(r17216)

These figures were arrived at by using [[|]]

== Advanced Tuning for Windows ==
This section has been included last for those who are technically capable.

=== Processor Scheduling (win32) ===
Windows enables the program running in the foreground to have heightened processor resources. If this option is switched to 'Backgrounf Processes', Tor appears to return consistently higher performance. This option can be found in 'My Computer | Right click | Properties | Advanced Tab | Performance Settings | Advanced tab again.

=== TCPOptimizer -  2K/XP's throughput (win32) ===
Windows XP has a self-tuning IP stack, but it can still benefit from a little help. Using the TCP Optimiser tool from above you can tune the MTU, RWIN, SACK OPTS (rfc 2038), and tcp1323opts controlling window scaling. The tool has one button optimise. This setting is sufficient to benefit from immediate increases to Tor throughput. To increase throughput further you can try experimenting with lower values of the IP TTL (Time To Live). Values as low as 32 will work and result in improved performance. Also try experimenting with smaller TCPWindowSizes. This setting is automatically adjusted when you move the slider marked 'Connection Speed' of the TCPOptimizer  tool.

You can view your connection parameters and their effects on your connections, both with Tor and without Tor, at [[|DSLReports]]

=== Event ID 4226 Patcher - Remove the limit on TCP connection attempts XP SP2 (win32) ===
[[|Remove the limit on TCP connection attempts]] has an interesting article detailing this restriction introduced in XP SP2. Microsoft have restricted the amount of half-open TCP/IP connections with the proviso that it would reduce the pace that worms spread. As noted by SpeedGuide, internet worms spread isotropically (multi-directionally) and so their infecton rate is exponential. As such, placing a constant (limit) on the rate of connection creation for every computer running XP SP2 will slow the rate of worms spreading (for that group of computers) but not by much. Consider the population of humans on the planet. Its over ~6 billion.

Supposing all these people are running Windows XP SP2, with rate limited half-open connections. Rate limiting is set to 10 half-open connections per second. To infect the entire population of computers would take: We are assuming optimum forward infection here. In the first second we have infected 10 machines. The 2nd second to elapse will cause (10 x 10) + 10 = 110 computers to be infected. The 3rd second to elapse would cause:

 . ( (10 x 10) x 10 ) + (10 * 10) + 10 =  1110 computers to be infected. So the number of computers infected for every second that elapses is : computers infected = ~ 10 ^ elapsedSeconds
In 12 seconds, we would have 10 ^ 12 = 1 billion computers infected. Full infection occurs before 13 seconds have elapsed !

This is all skewed by network topologies and routing algorithms, but they would affect a non-limited network in an identical manner. So the affect is a theoretical maximum of 13 seconds of additional notice to act against the worm. To all intents and purposes, this is useless.

Of much more interest is the effect on ANY network that relies on many open connections, such as Tor and a host of P2P applications. The effect here is a slow down of communications, with the limit acting as the catalyst.

Use the Event ID 4226 Patcher to mitigate against this.

The following parameters must also satisfy the formula below for optimal performance.

''Windows:RWIN >= Privoxy:buffer >= ( Tor:ConstrainedSockSize default = 262KB )''

||Windows||in TCP Optimizer||
|| -- ||-- ||
||RWIN|| TCPOptimizer max setting||
||MTU|| 576, 1500 or 1492 - try all 3 settings||
||Window Scaling|| On||
||TcpIP TTL || 128 (hops)||
||LANBufferSize || 65535 (bytes)||
Reducing DNS caching time reduces the risk of an invalid DNS resolve, given Tor servers may be operating in a DHCP environment that updates the IP each time the network connects.
||Windows||registry:TCPIP service||
|| -- || -- ||
||DNS Cache || 36000(seconds (10 hours)) set in registry by hand, HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\Dnscache\Parameters||
Privoxy is set to be a 'straight-through' proxy server, with the toggle switch. Its buffer is reduced to below that of RWIN. This is because RWIN represents the largest TCP receive window. Its value is chosen to be above Tor default socks size = 252KB
|| -- || -- ||
||Privoxy:buffer || 265 (KB)||
||toggle || 0||