715714 - Adapt cache based on io rates

Status: RESOLVED WONTFIX

(Core :: Networking: Cache, defect, P3)

Description

[(dormant account)]

On a lenovo x120e, deleting the cache directory takes about an hour. I don't remember the exact cache size, it was around 1gb(120gb of disk free).
We should restrict the size of cache on devices with extremely slow disks.
Cache performance also changes depending on whether other stuff is running on the computer. We should detect if cache entries are taking >100ms on average and disable disk cache completely in those cases.
This laptop consistently takes 500-800ms for HTTP_PAGE_CACHE_READ_TIME and is worse when other disk activity is present. Similar conditions happen on android phones.

Comment 1

[(dormant account)]

here is comment from my blog that shows another weird case of a need to adapt our cache to disk speed:

I have to limit my network cache to 100MB, otherwise shutdown times (and unclean startup times) explode beyond all belief.

I have my profile directory (including the cache) encrypted on Windows using EFS. Shutting down Firefox takes up to a minute, and after a process terminate a startup takes up to a minute.

It used to be beyond ridiculous before I reduced the network cache size. I mean, seriously ridiculous, I'd terminate Firefox while it was still starting up because I thought it was crashed.

Note: EFS is single-threaded on Windows. So even though I have 8 cores it could be using, it only uses one. Also running on SSD with 12GB ram.

Comment 2

[Josh Aas]

We'll be considering something like this as part of our Q1 cache performance work. Assigning to Nick Hurley as he's leading that effort.

Comment 5

[Firefox Bug Husbandry Bot]

Bulk change to priority: https://bugzilla.mozilla.org/show_bug.cgi?id=1399258

Comment 6

[Firefox Bug Husbandry Bot]

Bulk change to priority: https://bugzilla.mozilla.org/show_bug.cgi?id=1399258

Comment 7

[Michal Novotny [:michal]]

This shouldn't be problem anymore. Cache v2 doesn't have to remove the whole cache after a crash and RCWN should solve problems with very slow storages in general.