Hello, I'm doing research on improving Tor security via path selection. This work is with Ryan Wails, Aaron Johnson, Prateek Mittal, and Olivier Pereira. In our work, we ran into load-balancing problems that affected our experiments. We identified the problem to originate in the specification and implementation of Proposal 271, "Another algorithm for guard selection", which was implemented in tor-0.3.0.1-alpha ~2 years ago and is thus now in tor stable. In short, Prop 271 causes guards to be selected with probabilities different than their weights due to the way it samples many guards and then chooses primary guards from that sample. We are suggesting a straightforward fix to the problem, which is, roughly speaking, to choose primary guards in the order in which they were sampled. We have created a patch implementing this fix for the case affecting our experiments, which would improve the current situation. We are further suggesting that Tor apply the technique throughout the guard-selection logic. In more detail, Prop 271 chooses guards via a multi-step process:   1. It chooses 20 distinct guards (and sometimes more) by sampling without replacement with probability proportional to consensus weight.   2. It produces two subsets of the sample: (1) "filtered" guards, which are guards that satisfy various torrc constraints, and (2) "confirmed" guards, which are guards through which a circuit has been constructed.   3. The "primary" guards (i.e. the actual guards used for circuits) are chosen from the confirmed and filtered subsets. I'm ignoring the additional "usable" subsets for clarity. This description is based on Section 4.6 of the specification (https://gitweb.torproject.org/torspec.git/tree/guard-spec.txt). The primary guards are selected uniformly at random from the filtered guards when no confirmed guards exist. No confirmed guards appear to exist until some primary guards have been selected, and so when Tor is started the first time the primary guards always come only from the filtered set. The uniformly-random selection causes a bias in primary-guard selection away from consensus weights and towards a more uniform selection of guards. As just an example of the problem, if there were only 20 guards in the network, the sampled set would be all guards and primary guard selection would be entirely uniformly random, ignoring weights entirely. This bias is worse the larger the sampled set is relative to the entire set of guards, and it has a significant effect on Tor simulations, which are typically on smaller networks. We believe that this issue has only a limited effect on Tor currently due to the relatively large number of guards. However, the issue is potentially worse when there exist confirmed guards. In this case, primary guards are chosen from the intersection of confirmed and filtered guards in the order they were added to the confirmed set. That order seems to be the order in which circuits were successfully created through the guards, which would thus be both somewhat non-deterministic (because it depends on when a circuit successfully finishes) and random (because circuits are only attempted through primary guards, which are initially uniformly-random filtered guards). The non-determinism especially could yield guard selection probabilities that deviate quite a bit from the consensus weights. This design has both performance and security implications. It potentially reduces the performance of the Tor network by making the load unbalanced. It also affects the correctness of performance analyses done with Shadow simulations, and the error is likely much larger due to the smaller size of simulation networks, which commonly include 100 or fewer guards. The design also reduces Tor's security by increasing the number of clients that an adversary running small relays can observe. In addition, an adversary has to wait less time than it should after it starts a malicious guard to be chosen by a client. This weakness occurs because the malicious guard only needs to enter the sampled list to have a chance to be chosen as primary, rather than having to wait until all previously-sampled guards have already expired. We propose a solution that fits well within the existing guard-selection algorithm. Our solution is to select primary guards in the order they were sampled. This ordering should be applied after the filtering and/or confirmed guard sets are constructed as normal. That is, primary guards should be selected from the filtered guards (if no guards are both confirmed and filtered) or from the set of confirmed and filtered guards (if such guards exist) in the order they were initially sampled. This solution guarantees that each primary guard is selected (without replacement) from all guards with a probability that is proportional to its consensus weight. We include a patch that applies this ordering only when primary guards are selected from filtered guards. This fixes the problem in Shadow simulations because confirmed guards never exist when primary guards are being selected. However, to solve all issues in the real Tor network, we recommend that you apply similar logic to the case that primary guards are selected from confirmed guards. Note that the issue of sampling skewing guard selection seems to have been raised in proposal discussions, which is documented in the proposal (although the proposed solution is less practical):

[Paul Syverson in a conversation at the Wilmington Meeting 2017 says that we should look into how we're doing this sampling. Essentially, his concern is that, since we are sampling by bandwidth at first (when we choose the `sampled` set), then later there is another bias—when trying to build circuits (and hence marking guards as confirmed) we select those which completed a usable circuit first (and hence have the lowest latency)—that this sort of "doubly skewed" selection may "snub" some low-consensus-weight guards and leave them unused completely.  Thus the issue is primarily that we're not allocating network resources efficiently.  Mine and Nick's guard algorithm simulation code never checked what percentage of possible guards the algorithm reasonably allowed clients to use; this would be an interesting thing to check in simulation at some point.  If it does turn out to be a problem, Paul's intuition for a fix is to select uniformly at random to obtain the `sampled` set, then weight by bandwidth when trying to build circuits and marking guards as confirmed. —isis]

Best, Florentin