Evaluating QUIC Performance Over Web, Cloud Storage, and Video Workloads
Tanya Shreedhar, Rohit Panda, Sergey Podanev, Vaibhav Bajpai
Abstract
QUIC was launched in 2013 with a goal to provide reliable, connection-oriented and end-to-end encrypted transport and is recently standardized in May 2021 by the Internet Engineering Task Force (IETF). This work evaluates QUIC performance over the web, cloud storage, and video workloads and compares them to traditional TLS/TCP. To this end, we have designed tests ( <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">quic_perf</monospace> , <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">tls_perf</monospace> and <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">video</monospace> ) and conducted measurements from 2018 – 2021 using multiple vantage points: an educational network, a high-bandwidth low-RTT residential link in Germany and a low-bandwidth high-RTT residential link in India. We target Alexa Top-1M for web workloads and probe them towards the support for QUIC, TLS 1.2 and TLS 1.3. By measuring <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\lt 5.7$ </tex-math></inline-formula> K websites that support QUIC, we observe that QUIC has up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\approx 140$ </tex-math></inline-formula> % lower mean connection times than TLS 1.2/1.3 over TCP for low-bandwidth and high-RTT networks. When comparing different versions of QUIC, we observe that IETF QUIC connection times are slightly better than different versions ( <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q050</monospace> , <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q046</monospace> , <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q044</monospace> , <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q043</monospace> , <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q039</monospace> and <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Q035</monospace> ) of gQUIC. For cloud storage workloads, we observe that TLS 1.2 over TCP exhibits higher throughput for larger file sizes (>20 MB up to 2 GB), while QUIC exhibits higher throughput for smaller file sizes ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\leq 20$ </tex-math></inline-formula> MB) while downloading files from Google Drive. At the same time, QUIC has much higher CPU utilization than TLS 1.2 over TCP, almost double while downloading a large file (200 MB) from Google Drive due to in-kernel optimizations that benefit TCP. For video workloads, we observe that QUIC is 534 ms faster than TLS 1.2 over TCP from India (406 ms from Germany) in establishing a connection to YouTube media servers. Although we witness that (similar to cloud storage workloads) the overall download rate is higher over TLS, QUIC still tends to depict better video content delivery with reduced stall events and up to 50% lower stall durations due to its lower latency overheads. To support reproducibility, the developed tests and the collected data are made publicly available to the community.