2016 USENIX Annual Technical Conference (USENIX ATC '16)

A recent NSDI paper [1] reported that increasing the cache hit ratio for an HTTP proxy from 22% to 32% improved median page load time (PLT) formobile clients by less than 2%. We argue that there are two main causes for this weak improvement: objects on the critical path are often not cached, and the limited computational power of mobile devices causes computational delays to comprise a large portion of the critical path.

Both of these factors were, in fact, outlined by a previous analysis of desktop web performance [2]. However, we (as the authors of the HTTP proxy [1]) did not properly understand the analysis and could have saved ourselves substantial engineering costs ifwe had. We therefore argue for the need to highlight this prior analysis, and extend the analysis to include mobile devices with slow CPUs, precise cache hit ratios, and a controlled reproduction of the HTTP proxy caching results [1]. In the extreme case of a perfect cache hit ratio, desktop page load times are improved notably by 34% compared to no caching, but mobile page load times only improve by 13% in the median case. We extract a back-of-envelope performance model from these results to help understand their underlying causes.