1216286 - Find a benchmark which can be used for tuning GC settings.

Status: RESOLVED INACTIVE

(Core :: JavaScript: GC, defect)

Description

[Nicolas B. Pierron [:nbp]]

Currently, on B2G, the GC settings are tuned for an Unagi which is no longer representative of the devices that we have today.  To improve this state, we probably need better GC benchmarks for tuning the results.

For tuning the GC settings, we need a benchmark which can be used within the browser, as-if this was a standard web-page.  As opposed to normal web-pages, I think it makes sense to expose additional function to ensure that the garbage collection stats only account for the one compartment that we are testing, even if this means leaking in the rest of the browser.

Previously, I used a modified version of Bill benchmark for incremental GC, with a small workload (Bug 898556 comment 23) but this benchmark was extremlly noisy on the Unagi.  Terrence suggested using "devtools/gc-ubench" (Bug 1125398), but probably not for tuning GC performances.

So, the questions of this bug are:
 - Is there already a benchmark which is good enough to represent the web?
 - Could we make one out of the current set of gc-ubench by using weight on each benchmark result to give a score which is representative of the web?
 - Can we reduce the noise by having special builds which exposed internals of the GC?

Comment 1

[Nicolas B. Pierron [:nbp]]

This is no longer a concern for checking B2G.