Upgrading jemalloc
Paul Bone
jemalloc and Firefox
Most programs use the system-provided memory allocator.
malloc()
free()
realloc()
memalign()
calloc()
valloc()
Critical for performance, memory footprint and security of a computer program.
jemalloc and Firefox
- Forked from jemalloc 2
- In 2007
- Probably selected for good reasons, in 2007.
- Not great at multithreaded workloads, (multicore processors released in 2006)
- jemalloc is now version 5.3.0
Could we upgrade it?
Why don't we just?
Upgrade jemalloc
Upgrade jemalloc
In June 2017 I heard a converstation about a prior effort to do this.
Upgrade jemalloc
glandium: [We] tried jemalloc 3 and 4. ... the gist of it is that RSS was something like +50% while performance was slightly better, but tweaking to get less memory overhead made the performance not so good anymore.
Upgrade jemalloc
jesup tested jemalloc 5 in 2023.
- 4.45% faster (sp3)
- 300% memory usage (suspect bad result)
- Worth forther checking?
- TODO: compare other features including security.
Why don't we just
Use the system allocator
System allocator
December 2022
System allocator
| jemalloc | ptmalloc | increase | |
|---|---|---|---|
| resident-peak | 479MB | 552MB | 115% |
| resident | 202MB | 504MB | 250% |
Retested in 2024
pssst. Wait until the end of the presentation...
Missing features
- Poisioning
- Arenas
- API for releasing memory to the system
- System support varies by platform
Why don't we just?
Replace jemalloc
Replace it with mimalloc
jesup tested mimalloc in 2023
- Lock-free implementation
- Thread local free-lists
- Few branches on fast-path
- 3-4% performance win!
Replace it with mimalloc
jesup tested mimalloc in 2023
- 0.25 - 1.5% with security features enabled
- 38% additional memory usage
- No arenas
Chrome uses partition alloc
The implementation is too tightly tied to Chrome's infrastructure to test with Firefox.
Anything else?
A lot of allocators gave a ~5% performance win (at the time) but added a lot more memory overhead.
Lets try the hard way
We'll continue to maintain and improve mozjemalloc.
Prior improvments in mozjemalloc
- Guard pages around chunk payload.
- Reduced fragmentation for allocations between 512 and 4,096 bytes.
- Poisoning the first 256 botes of an object only.
Prior improvments in mozjemalloc
- Lock free access for DOM.
- Reduced lock contention by moving slow operations outside of locks
- VirtualAlloc calls amortised - without affecting memory usage
mozjemalloc
2 years ago disabling jemalloc yeilded a 5% perf gain.
Today jemalloc is 2.5% faster than the system allocator. And still consumes a lot less memory.
Future
- Unlock for purging memory (WIP).
- More control of when memory is purged (hooks to scheduler).
- Thread local caches.
- Batch allocation/free.
- More...