neighbor_hash_map

---

This repo contains codes and steps necessary to reproduce the artifacts for research paper titled CIKM'24: An Enhanced Batch Query Architecture in Real-time Recommendation.

Setting up the hardware

We conducted tests in the paper using Intel(R) Xeon(R) Gold 5218 CPUs. In fact, any hardware that supports AVX512 can be used for reproduction, but the results may not necessarily match the results presented in the paper.

Software Requirements

• Linux x86_64 kernel >= 4.9.0
• Hugepage (2 MiB) supported and has been configured, for example: echo 8192 > /proc/sys/vm/nr_hugepages
• Clang >= 16.0.6 or GCC >= 11.4.0
• Bazel >= 6.3.0
• Tested on Ubuntu 22.04

Build benchmark

git clone https://github.com/slow-steppers/NeighborHash
cd NeighborHash

When using clang

bazel build --config=clang --config=neighbor_hugepage --config=neighbor_simd //testing:all

Or, when using gcc

bazel build --config=neighbor_hugepage --config=neighbor_simd //testing:all

Run benchmarks

Our benchmark implementation utilizes the google benchmark framework, which allows selecting benchmarks to run using regular expressions. The regular expression for running random access is: BM_RandomAccess<.*HashMapType.*>/DS/SQR/LF. For instance, to run benchmarks for all scalar hashmaps with a 1M DatasetSize, SQR=90, and LF=80, execute the following code:

numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<.*Scalar.*uniform>/1048576/90/79"

We have integrated the following components into our benchmark:

• NeighborHash
• std::unordered_map
• LinearProbing
• BucketingSIMD_16x16
• ankerl::unordered_dense::map
• absl::flat_hash_map
• ska::bytell_hash_map
• emhash7::HashMap
• tsl::hopscotch_map
• Native array (Random Access)

For ease of comparison, we have categorized each hashmap into the following classes: Scalar, IntraVec, Vec, MultiThreading. The candidate values for each component of the expression are as follows:

• SQR: 30/50/90/100
• LF: 79
• DS: 1024/16384/131072/1048576/2097152/16777216/67108864/134217728
• HashMapType: Scalar/Vec/IntraVec/MultiThreading/QFGO

The following presents the code for reproducing the results from the paper:

# all scalar hashmaps under all dataset size with SQR=90,LF=80
numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<.*Scalar.*>/.*/90/79"

# Neighbor/LinearProbing/RandomAccess with AMAC under all dataset size with SQR=90,LF=80
numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<(Neighbor|LinearProbing|Array).*uniform.*AMAC.*>/.*/90/79$"

# all intra-vectorization hashmaps under all dataset size with SQR=90,LF=80
numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<.*IntraVec.*>/.*/90/79"

# all vectorization hashmaps under all dataset size with SQR=90,LF=80
numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<.*Vec.*>/.*/90/79"

# NeighborHash with QFGO under all dataset size with SQR=90,LF=80
numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/hash_map_benchmark --benchmark_filter="BM_RandomAccess<.*QFGO.*>/.*/90/79"

To perform Mixed insertions and lookups, use the following code:

numactl --membind=0 --cpunodebind=0 ./bazel-bin/testing/mixed_multi_threading