On Fri, Sep 30, 2016 at 08:38:52AM +0100, Pablo de Lara wrote: > This patchset improves lookup performance on the current hash library > by changing the existing lookup bulk pipeline, with an improved pipeline, > based on a loop-and-jump model, instead of the current 4-stage 2-entry > pipeline. > Also, x86 vectorized intrinsics are used to improve performance when > comparing signatures. > > First patch reorganizes the order of the hash structure. > The structure takes more than one 64-byte cache line, but not all > the fields are used in the lookup operation (the most common operation). > Therefore, all these fields have been moved to the first part of the > structure, > so they all fit in one cache line, improving slightly the performance in some > scenarios. > > Second patch modifies the order of the bucket structure. > Currently, the buckets store all the signatures together (current and > alternative). > In order to be able to perform a vectorized signature comparison, > all current signatures have to be together, so the order of the bucket has > been changed, > having separated all the current signatures from the alternative signatures. > > Third patch introduces x86 vectorized intrinsics. > When performing a lookup bulk operation, all current signatures in a bucket > are compared against the signature of the key being looked up. > Now that they all are together, a vectorized comparison can be performed, > which takes less instructions to be carried out. > In case of having a machine with AVX2, number of entries per bucket are > increased from 4 to 8, as AVX2 allows comparing two 256-bit values, with > 8x32-bit integers, > which are the 8 signatures on the bucket. > > Fourth (and last) patch modifies the current pipeline of the lookup bulk > function. > The new pipeline is based on a loop-and-jump model. The two key improvements > are: > > - Better prefetching: in this case, first 4 keys to be looked up are > prefetched, > and after that, the rest of the keys are prefetched at the time the > calculation > of the signatures are being performed. This gives more time for the CPU to > prefetch the data requesting before actually need it, which result in less > cache misses and therefore, higher throughput. > > - Lower performance penalty when using fallback: the lookup bulk algorithm > assumes that most times there will not be a collision in a bucket, but it > might > happen that two or more signatures are equal, which means that more than one > key comparison might be necessary. In that case, only the key of the first > hit is prefetched, > like in the current implementation. The difference now is that if this > comparison > results in a miss, the information of the other keys to be compared has > been stored, > unlike the current implementation, which needs to perform an entire simple > lookup again. > > Changes in v4: > - Reordered hash structure, so alt signature is at the start > of the next cache line, and explain in the commit message > why it has been moved > - Reordered hash structure, so name field is on top of the structure, > leaving all the fields used in lookup in the next cache line > (instead of the first cache line) > > Changes in v3: > - Corrected the cover letter (wrong number of patches) > > Changes in v2: > - Increased entries per bucket from 4 to 8 for all cases, > so it is not architecture dependent any longer. > - Replaced compile-time signature comparison function election > with run-time election, so best optimization available > will be used from a single binary. > - Reordered the hash structure, so all the fields used by lookup > are in the same cache line (first). > > Byron Marohn (3): > hash: reorganize bucket structure > hash: add vectorized comparison > hash: modify lookup bulk pipeline >
Hi, Firstly, checkpatches is reporting some style errors in these patches. Secondly, when I run the "hash_multiwriter_autotest" I get what I assume to be an error after applying this patchset. Before this set is applied, running that test shows the cycles per insert with/without lock elision. Now, though I'm getting an error about a key being dropped or failing to insert in the lock elision case, e.g. Core #2 inserting 1572864: 0 - 1,572,864 key 1497087 is lost 1 key lost I've run the test a number of times, and there is a single key lost each time. Please check on this, is it expected or is it a problem? Thanks, /Bruce
