In June, we announced a wide-scale post-quantum experiment with Google. We implemented two post-quantum (i.e., not yet known to be broken by quantum computers) key exchanges, integrated them into our TLS stack and deployed the implementation on our edge servers and in Chrome Canary clients. The goal of the experiment was to evaluate the performance and feasibility of deployment in TLS of two post-quantum key agreement ciphers.
In our previous blog post on post-quantum cryptography, we described differences between those two ciphers in detail. In case you didn’t have a chance to read it, we include a quick recap here. One characteristic of post-quantum key exchange algorithms is that the public keys are much larger than those used by "classical" algorithms. This will have an impact on the duration of the TLS handshake. For our experiment, we chose two algorithms: isogeny-based SIKE and lattice-based HRSS. The former has short key sizes (~330 bytes) but has a high computational cost; the latter has larger key sizes (~1100 bytes), but is a few orders of magnitude faster.
During NIST’s Second PQC Standardization Conference, Nick Sullivan presented our approach to this experiment and some initial results. Quite accurately, he compared NTRU-HRSS to an ostrich and SIKE to a turkey—one is big and fast and the other is small and slow.
Setup & Execution
We based our experiment on TLS 1.3. Cloudflare operated the server-side TLS connections and Google Chrome (Canary and Dev builds) represented the client side of the experiment. We enabled both CECPQ2 (HRSS + X25519) and CECPQ2b (SIKE/p434 + X25519) key-agreement algorithms on all TLS-terminating edge servers. Since the post-quantum algorithms are considered experimental, the X25519 key exchange serves as a fallback to ensure the classical security of the connection.
Clients participating in the experiment were split into 3 groups—those who initiated TLS handshake with post-quantum CECPQ2, CECPQ2b or non post-quantum X25519 public keys. Each group represented approximately one third of the Chrome Canary population participating in the experiment.
In order to distinguish between clients participating in or excluded from the experiment, we added a custom extension to the TLS handshake. It worked as a simple flag sent by clients and echoed back by Cloudflare edge servers. This allowed us to measure the duration of TLS handshakes only for clients participating in the experiment.
For each connection, we collected telemetry metrics. The most important metric was a TLS server-side handshake duration defined as the time between receiving the Client Hello and Client Finished messages. The diagram below shows details of what was measured and how post-quantum key exchange was integrated with TLS 1.3.
The experiment ran for 53 days in total, between August and October. During this time we collected millions of data samples, representing 5% of (anonymized) TLS connections that contained the extension signaling that the client was part of the experiment. We carried out the experiment in two phases.
In the first phase of the experiment, each client was assigned to use one of the three key exchange groups, and each client offered the same key exchange group for every connection. We collected over 10 million records over 40 days.
In the second phase of the experiment, client behavior was modified so that each client randomly chose which key exchange group to offer for each new connection, allowing us to directly compare the performance of each algorithm on a per-client basis. Data collection for this phase lasted 13 days and we collected 270 thousand records.
We now describe our server-side measurement results. Client-side results are described at https://www.imperialviolet.org/2019/10/30/pqsivssl.html.
What did we find?
The primary metric we collected for each connection was the server-side handshake duration. The below histograms show handshake duration timings for all client measurements gathered in the first phase of the experiment, as well as breakdowns into the top five operating systems. The operating system breakdowns shown are restricted to only desktop/laptop devices except for Android, which consists of only mobile devices.
It’s clear from the above plots that for most clients, CECPQ2b performs worse than CECPQ2 and CONTROL. Thus, the small key size of CECPQ2b does not make up for its large computational cost—the ostrich outpaces the turkey.
Digging a little deeper
This means we’re done, right? Not quite. We are interested in determining if there are any populations of TLS clients for which CECPQ2b consistency outperforms CECPQ2. This requires taking a closer look at the long tail of handshake durations. The below plots show cumulative distribution functions (CDFs) of handshake timings zoomed in on the 80th percentile (e.g., showing the top 20% of slowest handshakes).
Here, we start to see something interesting. For Android, Linux, and Windows devices, there is a crossover point where CECPQ2b actually starts to outperform CECPQ2 (Android: ~94th percentile, Linux: ~92nd percentile, Windows: ~95th percentile). macOS and ChromeOS do not appear to have these crossover points.
These effects are small but statistically significant in some cases. The below table shows approximate 95% confidence intervals for the 50th (median), 95th, and 99th percentiles of handshake durations for each key exchange group and device type, calculated using Maritz-Jarrett estimators. The numbers within square brackets give the lower and upper bounds on our estimates for each percentile of the “true” distribution of handshake durations based on the samples collected in the experiment. For example, with a 95% confidence level we can say that the 99th percentile of handshake durations for CECPQ2 on Android devices lies between 4057ms and 4478ms, while the 99th percentile for CECPQ2b lies between 3276ms and 3646ms. Since the intervals do not overlap, we say that with statistical significance, the experiment indicates that CECPQ2b performs better than CECPQ2 for the slowest 1% of Android connections. Configurations where CECPQ2 or CECPQ2b outperforms the other with statistical significance are marked with green in the table.
A second phase of the experiment directly examined the performance of each key exchange algorithm for individual clients, where a client is defined to be a unique (anonymized) IP address and user agent pair. Instead of choosing a single key exchange algorithm for the duration of the experiment, clients randomly selected one of the experiment configurations for each new connection. Although the duration and sample size were limited for this phase of the experiment, we collected at least three handshake measurements for each group configuration from 3900 unique clients.
The plot below shows for each of these clients the difference in latency between CECPQ2 and CECPQ2b, taking the minimum latency sample for each key exchange group as the representative value. The CDF plot shows that for 80% of clients, CECPQ2 outperformed or matched CECPQ2b, and for 99% of clients, the latency gap remained within 70ms. At a high level, this indicates that very few clients performed significantly worse with CECPQ2 over CECPQ2b.
Do other factors impact the latency gap?
We looked at a number of other factors—including session resumption, IP version, and network location—to see if they impacted the latency gap between CECPQ2 and CECPQ2b. These factors impacted the overall handshake latency, but we did not find that any made a significant impact on the latency gap between post-quantum ciphers. We share some interesting observations from this analysis below.
Approximately 53% of all connections in the experiment were completed with TLS handshake resumption. However, the percentage of resumed connections varied significantly based on the device configuration. Connections from mobile devices were only resumed ~25% of the time, while between 40% and 70% of connections from laptop/desktop devices were resumed. Additionally, resumption provided between a 30% and 50% speedup for all device types.
We also examined the impact of IP version on handshake latency. Only 12.5% of the connections in the experiment used IPv6. These connections were 20-40% faster than IPv4 connections for desktop/laptop devices, but ~15% slower for mobile devices. This could be an artifact of IPv6 being generally deployed on newer devices with faster processors. For Android, the experiment was only run on devices with more modern processors, which perhaps eliminated the bias.
The slow connections making up the long tail of handshake durations were not isolated to a few countries, Autonomous Systems (ASes), or subnets, but originated from a globally diverse set of clients. We did not find a correlation between the relative performance of the two post-quantum key exchange algorithms based on these factors.
We found that CECPQ2 (the ostrich) outperformed CECPQ2 (the turkey) for the majority of connections in the experiment, indicating that fast algorithms with large keys may be more suitable for TLS than slow algorithms with small keys. However, we observed the opposite—that CECPQ2b outperformed CECPQ2—for the slowest connections on some devices, including Windows computers and Android mobile devices. One possible explanation for this is packet fragmentation and packet loss. The maximum size of TCP packets that can be sent across a network is limited by the maximum transmission unit (MTU) of the network path, which is often ~1400 bytes. During the TLS handshake the server responds to the client with its public key and ciphertext, the combined size of which exceeds the MTU, so it is likely that handshake messages must be split across multiple TCP packets. This increases the risk of lost packets and delays due to retransmission. A repeat of this experiment that includes collection of fine-grained TCP telemetry could confirm this hypothesis.
A somewhat surprising result of this experiment is just how fast HRSS performs for the majority of connections. Recall that the CECPQ2 cipher performs key exchange operations for both X25519 and HRSS, but the additional overhead of HRSS is barely noticeable. Comparing benchmark results, we can see that HRSS will be faster than X25519 on the server side and slower on the client side.
In our design, the client side performs two operations—key generation and KEM decapsulation. Looking at those two operations we can see that the key generation is a bottleneck here.
Key generation: 3553.5 [ops/sec] KEM decapsulation: 17186.7 [ops/sec]
In algorithms with quotient-style keys (like NTRU), the key generation algorithm performs an inversion in the quotient ring—an operation that is quite computationally expensive. Alternatively, a TLS implementation could generate ephemeral keys ahead of time in order to speed up key exchange. There are several other lattice-based key exchange candidates that may be worth experimenting with in the context of TLS key exchange, which are based on different underlying principles than the HRSS construction. These candidates have similar key sizes and faster key generation algorithms, but have their own drawbacks. For now, HRSS looks like the more promising algorithm for use in TLS.
In the case of SIKE, we implemented the most recent version of the algorithm, and instantiated it with the most performance-efficient parameter set for our experiment. The algorithm is computationally expensive, so we were required to use assembly to optimize it. In order to ensure best performance on Intel, most performance-critical operations have two different implementations; the library detects CPU capabilities and uses faster instructions if available, but otherwise falls back to a slightly slower generic implementation. We developed our own optimizations for 64-bit ARM CPUs. Nevertheless, our results show that SIKE incurred a significant overhead for every connection, especially on devices with weaker processors. It must be noted that high-performance isogeny-based public key cryptography is arguably much less developed than its lattice-based counterparts. Some ideas to develop this are floating around, and we hope to see performance improvements in the future.