Discussing with Fabrice, the wroker are consuming way too much memory to be efficient. B2G 1.3 (Gecko 28) is moving to used shared worker, which means that almost every application is going to use a worker at the end of 1.5. Discussing with terrence, this might either be a probe issue or the fact that we do not decommit the memory as we are supposed to do on a shrinking GC. Currently the attached profile show that each worker has 3 zones, one for the atoms compartment, one for the self-hosted code and one for the worker compartment, each of them are taking ~170 KB and correspond to half the memory consumed by the worker runtime.

This might be a good argument for lazy-parsing the self-hosted code. I didn't get any perf improvements out of that when I last tried (in bug 900292), so I didn't follow through and fixed all bugs, but I didn't really look into memory usage. Another thing to look into again is the function re-lazification in bug 886193.

(In reply to Till Schneidereit [:till] from comment #1) > This might be a good argument for lazy-parsing the self-hosted code. I > didn't get any perf improvements out of that when I last tried (in bug > 900292), so I didn't follow through and fixed all bugs, but I didn't really > look into memory usage. I don't think this is related to the self-hosted code yet, I would be surprised if it is because the 3 zones are exactly the same size and I would not expect the self-hosted code to fill all 3 zones the same way.

(In reply to Nicolas B. Pierron [:nbp] from comment #2) > I don't think this is related to the self-hosted code yet, I would be > surprised if it is because the 3 zones are exactly the same size and I would > not expect the self-hosted code to fill all 3 zones the same way. Mmh, that's a good point. It might be explained by most of the self-hosting compartment's memory being shared via the shared bytecode thing. Can you check how much memory is used by that? Self-hosted code should be responsible for almost all of it.

This bug is about memory consumption of JSRuntime in workers. Bug 864927 is about memory consumption of workers in general, plus it has much more discussion. There's no point having two levels of meta bugs, so I will dup this bug to that one.