Re: [PATCH] malloc: Use correct C11 atomics for fastbin

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi DJ,

> If testing shows that a larger default makes sense across a wide range
> of applications, then we can change the default.  I'm a bit opposed to
> changing anything because "it makes sense".  Back it up with data.

Larger values result in major speedups. Even a very conservative increase
to 50 speeds up xalancbmk up by 5% and omnetpp by 1.3%. A value of
2000 gives 17.5% and 6% respectively, but performance increases continue
well beyond that.

> tcache is wide but shallow, fastbins are narrow but deep (tcache caches
> much larger chunks than fastbins do) Benchmarks show they both provide a
> speed boost separately, and more so when combined.

If tcache was deeper, even if only for smaller sizes, it would make fastbins
completely redundant and with it all the complex consolidation. As a result,
there would be less fragmentation since adjacent free blocks will be merged
together more often.

>> Modern allocators are not only much faster than GLIBC in their default
>> settings
>
> Back this up with data please.  Tcache invalidated most online
> benchmarks.

Here is a recent paper that covers GLIBC with tcache - it still loses both on
performance and memory usage: https://arxiv.org/pdf/1905.01135.pdf 

>> but also have lower memory usage.
>
> This is a project we've got on our to-do list.

I don't believe you can fix that without a major rewrite.

Cheers,
Wilco