Memory usage reported by RSS is much higher than RocksDB MemoryUtil

[979357361]

we use RocksDB v9.4 with C API,
we want to control the total memory used by rocksdb, the WIKI told me there are mainly four parts contribute to RAM usage:
memtable / block cache / table reader / iterator pinned block.
we only use get and multi-get, so iterator can be filter out. Here is my setting:

1. key_size = 19B, val_sz = 22B, SST = 64MB,
2. memtable = 64MB, at most two memtables, that is 128MB at most
3. 10bit bloom filter per kv, for each 250M kv that is 310MB filter blocks
4. block_size = 4KB, for each 250M kv, there is 250M * (1+19+1+22) / 4K = 2750K bocks (ken_len, key, val_len, val), leading to 2750K * (19+4+1) = 65MB index blocks ( a key, a offset and size)
5. for each 250M kv, we reserved 450MB > (310MB + 65MB) RAM for table reader (or called meta blocks -- fd, filter blocks and index blocks)
6. for each 250M kv, we set hyper_block_cache = 440MB

We want rocksdb to consume <1GB RAM per 250M kv, we create rocksdb_memory_consumers_create, add block cache and DB to it, after feeding the DB 250M kv and warm up reading, we get from rocksdb_memory_consumers_create that the three usages are exactly what we thought:
rocksdb_approximate_memory_usage_get_mem_table_total < 64MB
rocksdb_approximate_memory_usage_get_mem_table_readers_total ≈ 440MB
rocksdb_approximate_memory_usage_get_cache_total ≈ 440MB

BUT, the RES in top command shows RocksDB consumed 1.5GB RAM. which is 50% higher than we thought, more over, when we get from rocksdb_memory_consumers_create the total memory is 8GB, the RES shows us 12GB RAM usage, also 50% higher

why the Memory usage reported by RSS is much higher than RocksDB MemoryUtil? Are our calculations wrong?

Any comments you may make are greatly appreciated, thanks you!

[jowlyzhang]

Do you use allow_mmap_read or allow_mmap_write, the mmapped memory is not accounted in those categories.
Looks like mmap_read is supposed to be used in combination of no compression + no block cache. So that probably is not the case here.

[979357361]

Do you use allow_mmap_read or allow_mmap_write, the mmapped memory is not accounted in those categories. Looks like mmap_read is supposed to be used in combination of no compression + no block cache. So that probably is not the case here.

thank you @jowlyzhang , I will have a try on mmap related options
we didn't change any mmap-related options, there is no compression but we did set a hyper clock cache,