Feature #8228
open[SDKs] [FUSE] Python SDK and arv-mount use Range requests when a caller requests part of a block that has been ejected from the cache
Description
Background¶
In crunch jobs (and elsewhere) it's easy to demand more random access than arv-mount's cache can handle (e.g., by reading from 100 files in order to do a merge sort). Currently, the only way to deal with this is to increase arv-mount's block cache: even without readahead/prefetch, this means 6.4GiB of RAM is needed to support a 100-way merge. If your cache is smaller, arv-mount can easily end up reading each 64 MiB block hundreds of times as it enters and leaves the cache. To make matters worse, in order to detect that arv-mount's cache is thrashing, you need to pay attention to crunchstat logs.
This story aims to reduce cache requirements such that a reasonable data cache size (like the default 256 MiB) is enough to support a reasonable number of concurrent readers (like 512).
There are two basic changes:- Make it possible to eject most of a 64 MiB block from the cache without ejecting all of it. If each of 10 processes is reading 64 KiB at a time, a 256 MiB cache is more effective if it holds 500 x 512 Kib chunks rather than 4 x 64 MiB chunks.
- Make it possible to retrieve part of a 64 MiB block from Keepstore without retrieving all of it. In a situation where we've already ejected a 64 MiB block from the cache to reclaim space, if we need part of that block again, ejecting another 64 MiB block from the cache to make room for this one again is just making the problem worse. And if we're not going to cache it, we shouldn't bother transferring it over the network or asking Keepstore to read it from disk. Instead, we should request a portion small enough to put in our cache.
We want to get these advantages without sacrificing our use of the MD5 hash to check data integrity. (In fact, it is more important to check data integrity when doing partial reads: Keepstore can't do it for us if it's not reading the whole block from disk.)
Implementation¶
In the Python SDK,KeepClient.get()
should
- have a
CACHE_CHUNK_SIZE
const = 524288 - accept a range argument (e.g.,
range=[1234,11234]
), and return only the specified byte range. - instead of inserting an entire block in the cache, chunk it into CACHE_CHUNK_SIZE-byte pieces and insert each chunk into the cache (the cache key will now be a tuple
(md5, byte_offset)
) - maintain a separate cache of chunk hashes: (md5, byte_offset) → chunk_md5 (capped at data_cache_cap_bytes÷32768 entries, which is approximately 1/1024 of the size of the data cache and will accept 16x)
- do a range request that returns all of the required pieces (i.e., the Range header boundaries sent to Keepstore will be a multiple of
CACHE_CHUNK_SIZE
, even if the boundaries in the range argument to get() aren't) - if the Keepstore service ignores the Range header and returns the entire block, discard the excess parts. (Sure, it might be better to cache them just in case they're needed, but that might also cause more cache churn and wreck performance, and it's a temporary measure to accommodate older services anyway.)
- fetch the whole block
- Insert/freshen all of the returned chunks in the data cache
- Insert/freshen all of the chunk MD5s in the chunk-hash cache
KeepClient.get_from_cache()
should also accept a range argument, and look up the appropriate entries in the data cache to put together a response.
ArvadosFile.readfrom
should pass a range argument to get_block_contents, and get_block_contents should pass that range argument along to self._keep.get()
and self._keep.get_from_cache()
.
ArvadosFile._BlockManager._block_prefetch_worker
should call self._keep.get(b, range=(0,1))
instead of self._keep.get(b)
to avoid evicting a lot of useful data from the cache unless that's necessary to fill the chunk-MD5 cache.
Related issues
Updated by Tom Clegg over 8 years ago
- Description updated (diff)
- Category set to SDKs
Updated by Brett Smith over 8 years ago
- Target version set to Arvados Future Sprints
Updated by Ward Vandewege over 3 years ago
- Target version deleted (
Arvados Future Sprints)