More backend work in the disk cache is demanding good tools for ongoing disk cache performance tracking. The Very Simple Backend implementation needs continuous A/B testing to show its implementation is advancing in speed. To track progress on this backend, and also to permit comparisons between alternative backends provided in net, two ongoing methodologies are proposed.
Proxy Backend with Replay
The Proxy Backend is a simple Disk Cache Backend
and Entry implementation that pass through to an underlying Entry and Backend, but recording a short log with parameter information & timing information to allow replay.
This log can then be replayed in a standalone replay application which takes the log, constructs a backend (perhaps a standard blockfile backend, a very simple backend
or a log structured backend), and performs the same operations at the same delays, all calls as if from the IO thread. The average latency of calls, as well as system load during the test can then permit A/B comparison between backends.
- Well suited to use on developer workstations.
- Very simple to collect logs and run.
- Very fast to run tests.
- Low level, includes very little noise from outside of the disk cache.
- Using logs with multiple versions of the same backend allows tracking progress over time of a particular backend.
- Sensitive to evictions: if two different backends have different eviction algorithms, the same operations on two backends can result in different logs. For instance an OpenEntry() on the foo backend could find an entry that then is Read/Written to when the same log played on a bar backend.
- Compares all low level operations equally: backend operations in the critical path of requests block launching requests. Other backend operations (like WriteData) almost never occur in the critical path, and so performance may impact rendering less. Without a full renderer, this impact is hard to measure.
- Does not track system resource consumption. Besides answering requests, the backend is consuming finite system resources (RAM, buffer cache, file handles, etc...), competing with the renderer for resources. This impact isn't very well measured in the replay benchmark.
browser_test with corpus
Starting up with a backend, the browser_test loads a large corpus of pages (either from a local server, or a web server simulating web latency, or possibly even the actual web), with an initially empty cache, and then runs either that same corpus or a second corpus immediately afterwards with the warm cache. This can be repeated using different backends, to permit A/B comparisons.
The browser test should introspect on UMA; outputs should include HttpCache.* benchmarks, as well as PLT.BeginToFirstPaint/PLT.BeginToFinish.
- Well suited to ongoing performance dashboards.
- Tracks a metric more closely connected to user experience.
- Allows comparison of backends with wildly different eviction behaviour.
- Noisier, since it includes much, much more than just the Chrome renderer.
- Slower to run.
- Because of ongoing renderer changes, comparisons of the same backend over time are problematic without a lot of patch cherry picking.