Arvados

The Keep client currently uses an in-memory cache to store blocks that have been loaded from Keep. The cache replacement policy implemented is Least Recently Used (LRU).

Given that the amount of memory in our machines is relatively small compared to the amount of data being pulled from Keep, we observe a significant amount of cache thrashing.

It would be useful if the Keep client had the ability to exploit the relatively large amount of disk available on our machines for caching. This would allow for clients with a larger cache, without requiring more memory. It could also be advantageous if this cache could be shared between multiple clients running on the same machine. The ability to use different cache replacement policies would be useful to reduce cache thrashing.