This is the mail archive of the
mailing list for the glibc project.
Re: [RFC/PoC] malloc: use wfcqueue to speed up remote frees
- From: Carlos O'Donell <carlos at redhat dot com>
- To: Eric Wong <normalperson at yhbt dot net>
- Cc: libc-alpha at sourceware dot org
- Date: Thu, 2 Aug 2018 17:38:19 -0400
- Subject: Re: [RFC/PoC] malloc: use wfcqueue to speed up remote frees
- References: <20180731084936.g4yw6wnvt677miti@dcvr> <firstname.lastname@example.org> <20180731231819.57xsqvdfdyfxrzy5@whir> <email@example.com> <20180801062352.rlrjqmsszntkzlfe@untitled> <firstname.lastname@example.org> <20180801092626.jrwyrojfye4avcis@whir>
On 08/01/2018 05:26 AM, Eric Wong wrote:
> Carlos O'Donell <email@example.com> wrote:
>> On 08/01/2018 02:23 AM, Eric Wong wrote:
>>> Carlos O'Donell <firstname.lastname@example.org> wrote:
>>>> On 07/31/2018 07:18 PM, Eric Wong wrote:
>>>>> Also, if I spawn a bunch of threads and get a bunch of
>>>>> arenas early in the program lifetime; and then only have few
>>>>> threads later, there can be a lot of idle arenas.
>>>> Yes. That is true. We don't coalesce arenas to match the thread
>>> Eep :< If contention can be avoided (which tcache seems to
>>> work well for), limiting arenas to CPU count seems desirable and
>>> worth trying.
>> In general it is not as bad as you think.
>> An arena is made up of a chain of heaps, each an mmap'd block, and
>> if we can manage to free an entire heap then we unmap the heap,
>> and if we're lucky we can manage to free down the entire arena
>> (_int_free -> large chunk / consolidate -> heap_trim -> shrink_heap).
>> So we might just end up with a large number of arena's that don't
>> have very much allocated at all, but are all on the arena free list
>> waiting for a thread to attach to them to reduce overall contention.
>> I agree that it would be *better* if we had one arena per CPU and
>> each thread could easily determine the CPU it was on (via a
>> restartable sequence) and then allocate CPU-local memory to work
>> with (the best you can do; ignoring NUMA effects).
> Thanks for the info on arenas. One problem for Ruby is we get
> many threads, and they create allocations of varying
> lifetimes. All this while malloc contention is rarely a
> problem in Ruby because of the global VM lock (GVL).
The allocations of varying lifetimes will make it impossible to free
down a heap from a heap-based allocator. This is a serious issue with
heap-based allocators, and it will impact the max RSS that you'll
need to reach steady state. It's not really a tractable problem I think,
I don't know how to deinterlace the the chunks which have differing
lifetimes. Your only chance is to take the existing large/small/fast
bin machinery and instead of mixing them in one heap, split them into
one smaller heap each, and see how that goes i.e. adopt size classes,
but keep it heap-based.
> Even without restartable sequences, I was wondering if lfstack
> (also in urcu) could even be used for sharing/distributing
> arenas between threads. This would require tcache to avoid
> retries on lfstack pop/push.
We absolutely need a better balancing across arenas, even now we don't
do any rebalancing based on load or attach count. We should. That
problem would go away if you just used restartable sequences to find
your cpu, map that to the local arena for the cpu and allocate there.
> Much less straighforward than using wfcqueue for frees with
> this patch, though :)
> Heh, a bit dirty, but #define-ing poll away seems to work :)
> diff --git a/malloc/malloc.c b/malloc/malloc.c
> index 40d61e45db..89e675c7a0 100644
> --- a/malloc/malloc.c
> +++ b/malloc/malloc.c
> @@ -247,6 +247,11 @@
> /* For SINGLE_THREAD_P. */
> #include <sysdep-cancel.h>
> +/* prevent wfcqueue.h from including poll.h and linking to it */
> +#include <poll.h>
> +#undef poll
> +#define poll(a,b,c) assert(0 && "should not be called")
Call __poll instead. That should fix the issue.
> #define _LGPL_SOURCE /* allows inlines */
> #include <urcu/wfcqueue.h>