How to increase the warmup cache for HNSW + MMAP + Local SSD disk

[xiaobingxia-at]

Hello, I'm using the HNSW + MMAP + Local SSD disk setup.
I observed that when doing the first search given specific collection, and corresponding filtering by partition key and clustering key, the first search would be 50-100ms. And following searches for same collection, partition key, clustering key will be <10ms.

But if the total data volume exceeds some level, when doing the intensive searches with a couple of threads, the search acceleration effect disappear. I assume that we run out of cache memory, and old entries got evicted from the cache.

Is there anyway to increase the warmup memory cache size for HNSW + MMAP + Local SSD setup?

Thanks!

[xiaofan-luan]

Hello, I'm using the HNSW + MMAP + Local SSD disk setup. I observed that when doing the first search given specific collection, and corresponding filtering by partition key and clustering key, the first search would be 50-100ms. And following searches for same collection, partition key, clustering key will be <10ms.

But if the total data volume exceeds some level, when doing the intensive searches with a couple of threads, the search acceleration effect disappear. I assume that we run out of cache memory, and old entries got evicted from the cache.

Is there anyway to increase the warmup memory cache size for HNSW + MMAP + Local SSD setup?

Thanks!

it is baed on your querynode memory limit.
Milvus always try to fully utilize the mmory of you pod with page cache

[xiaobingxia-at]

My pod has enough memory, per dashboard, only 50% of memory is being used. However, first groups of searches are still slow. I hope it could use more memory.