WaLSM is a workload-aware KV store for NVM-SSD hybrid storage. It is designed for reducing write amplifications on SSD with NVM, and auto index tuning on varying workloads.
Our code is on dev branch.
Currently, we only support building with Makefile. Use
make -j32 static_lib
to build the static library.
We use the following settings for testing performance:
opt.create_if_missing = true;
opt.use_direct_io_for_flush_and_compaction = true;
opt.use_direct_reads = true;
opt.compression = rocksdb::kNoCompression;
opt.compaction_style = rocksdb::kCompactionStyleUniversal;
opt.IncreaseParallelism(32);
opt.statistics = rocksdb::CreateDBStatistics();
opt.nvm_path = nvm_path; // use nvm path here
rocksdb::BlockBasedTableOptions block_based_options;
block_based_options.pin_top_level_index_and_filter = false;
block_based_options.pin_l0_filter_and_index_blocks_in_cache = false;
block_based_options.cache_index_and_filter_blocks_with_high_priority = false;
block_based_options.index_type = rocksdb::BlockBasedTableOptions::kTwoLevelIndexSearch;
block_based_options.partition_filters = true;
block_based_options.cache_index_and_filter_blocks = true;
block_based_options.filter_policy.reset(rocksdb::NewBloomFilterPolicy(10, false));
block_based_options.block_cache =
rocksdb::NewLRUCache(static_cast<size_t>(128 * 1024 * 1024));
opt.table_factory.reset(rocksdb::NewBlockBasedTableFactory(block_based_options));
opt.memtable_prefix_bloom_size_ratio = 0.02;Note that our compacting algorithm is based on Universal Compaction, using other RocksDB default compacting algorithm may cause unexpected behaviors.
