Once we are finding reftests via rel=match, we can then have multiple tests using the same reference URL. There's no reason to load the same reference URL hundreds (or even thousands, in the most common case!) of times. e.g. ref-filled-green-100px-square.xht is used >2k times; if we assume loading and doing the screenshot takes 25ms each time, then almost a minute is spent just taking screenshots of that URL. Actually caching screenshots probably also would significantly decrease the cost of importing css/CSS2 (Bug 209735), as CSS2 counts for a large percentage of all reftests in WPT. My initial inclination would be to cache per test runner worker (as this avoids having to synchronise caches).