Way back in time, I used delta encoding for storing posting list (inverted index for search index). I experimented with using GPUs for decoding the posting list. It turned out that, as another reply mentioned copying posting list from CPU memory to GPU memory was taking way too long. If posting list is static, it can be copied to GPU memory once. This will make the decoding faster. But still there is a bottle neck of copying the result back into CPU memory.
Nvidia's unified memory architecture may make it better as same memory can be shared between CPU and GPU.
While the results look impressive, I can't help but think "yeah but had you stored an absolute value every X deltas instead of just a stream of deltas, you would have had a perfectly scalable parallel decoding"
"We achieve 19.8 GB/s prefix sum throughput—1.8x faster than a naive implementation and 2.6x faster than FastPFoR"
"FastPFoR is well-established in both industry and academia. However, on our target platform (Graviton4, SIMDe-compiled) it benchmarks at only ~7.7 GB/s, beneath a naive scalar loop at ~10.8 GB/s."
I thought the first bit was a typo but it was correct; the naive approach was faster than a "better" method. Another demonstration of how actually benchmarking on the target platform is important!
6 comments
[ 3.3 ms ] story [ 23.3 ms ] threadNvidia's unified memory architecture may make it better as same memory can be shared between CPU and GPU.
"FastPFoR is well-established in both industry and academia. However, on our target platform (Graviton4, SIMDe-compiled) it benchmarks at only ~7.7 GB/s, beneath a naive scalar loop at ~10.8 GB/s."
I thought the first bit was a typo but it was correct; the naive approach was faster than a "better" method. Another demonstration of how actually benchmarking on the target platform is important!
Fun project btw, HPC is always interesting.