Re: SIMD accelerated crush_do_rule proof of concept

[Loic Dachary <loic@dachary.org>]

Hi Mark,

On 29/08/2016 15:15, Mark Nelson wrote:
> Anything we can do to help on the CPU usage front is a win IMHO, though I would be interested in seeing an example where we are spending a lot of time on crush in a real usage scenario?

When the ceph setcrushmap sends a new map to the monitor, it runs crushtool --test[1] and blocks until it's done. When the crushmap is large or the rules are complex it can timeout because the monitor won't let it run for longer than the mon lease time.

Cheers

[1] https://github.com/ceph/ceph/blob/master/src/crush/CrushTester.cc#L361

> Mark
> 
> On 08/29/2016 06:42 AM, Loic Dachary wrote:
>> Hi,
>>
>> TL;DR: crush_do_rule using SIMD goes twice faster, the implementation is straightforward and would help with crushmap validation, is there any reason not to do it ?
>>
>> When resolving a crush rule (crush_do_rule in mapper.c), the straw2 function (bucket_straw2_choose) calls the hashing function (crush_hash32_3) for each item in a bucket and keeps the best match. When a bucket has four items, the hash function can be run using SIMD instructions. Each item value is 32 bits and four can fit in a __m128i.
>>
>> I tried to inline the hash function when the conditions are right[1] and run a test to measure the difference.
>>
>> crushtool -o /tmp/t.map --num_osds 1024 --build node straw2 8 datacenter straw2 4 root straw2 0
>> time crushtool -i /tmp/t.map --show-bad-mappings --show-statistics --test --rule 0 --min-x 1 --max-x 2048000 --num-rep 4
>> rule 0 (replicated_ruleset), x = 1..2048000, numrep = 4..4
>> rule 0 (replicated_ruleset) num_rep 4 result size == 4:    2048000/2048000
>>
>> With SIMD
>>
>> real    0m10.433s
>> user    0m10.428s
>> sys    0m0.000s
>>
>> Without SIMD
>>
>> real    0m19.344s
>> user    0m19.340s
>> sys    0m0.004s
>>
>> Callgrind estimated cycles for each crush_do_rule are in the same range:
>>
>> rm crush.callgrind ; valgrind --tool=callgrind --callgrind-out-file=crush.callgrind crushtool -i /tmp/t.map --show-bad-mappings --show-statistics --test --rule 0 --min-x 1 --max-x 204800 --num-rep 4
>> kcachegrind crush.callgrind
>>
>> With SIMD    : crush_do_rule is estimated to use 21 205 cycles
>> Without SIMD : crush_do_rule is estimated to use 53 068 cycles
>>
>> This proof of concept relies on instructions that are available on all ARM & Intel processors, nothing complicated is going on. It is beneficial to crush maps that have more than four disks per host, more than four hosts per rack etc. It probably is a small win for an OSD or even a client. For crushmap validation it helps significantly since the MON are not able to run crushtool asynchronously and it needs to run within a few seconds (because it blocks the MON).
>>
>> The implementation is straightforward: it needs sub/xor/lshift/rshift. The only relatively tricky part is runtime / compile time detection of the SIMD instructions for both Intel and ARM processors. Luckily this has already been taken care of when integrating with the jerasure erasure code plugin.
>>
>> Is there any reason why it would not be good to implement this ?
>>
>> Cheers
>>
>> [1] https://github.com/dachary/ceph/commit/71ae4584d9ed57f70aad718d0ffe206a01e91fef
>>
> 

-- 
Loïc Dachary, Artisan Logiciel Libre