[neo-th]

Problem

Chunk 0 of every object went to the first drive (block_devices[chunk_idx % num_devices]). That has two effects:

• Objects that don't fill a whole stripe leave the last drives underused. A 14 MiB object in 4 MiB chunks puts 4 MiB on drives 1 to 3 and 2 MiB on drive 4.
• Objects under one chunk all land on the first drive, however many drives are in the pool.

Change

The start drive now comes from a hash of the object's URI.

• The hash is stable across restarts, so an append continues the object's pattern.
• Each chunk records its own drive, so data that's already stored stays readable.
• test_server_stage_write_and_commit_flow required chunk 0 on drive 0. It now checks that the object's two chunks use both drives.

Results

These are from 4 Micron 7450 Gen4 drives through LMCache's L2 benchmark, with 64 objects in flight and 32 workers. Each figure is the median of 3 runs in shuffled order, old and fixed in the same session.

Objects	Load before (MB/s)	Load after (MB/s)	Store before (MB/s)	Store after (MB/s)
2 MiB	6321	24609	4962	15390
14 MiB	20740	24884	16129	18549
16 MiB	23587	24646	18144	18759

At 14 MiB on 2 drives, load goes from 10526 to 12071 MB/s. fio measured 27.6 GB/s read on the same 4 drives at 4 MiB QD32.

Testing

All on this code or the same fix on the connector branch, on 4 Micron 7450 drives:

• Old vs fixed: 3 runs of each setup in shuffled order, in one session.
• Upgrade: from today's main through the Python client, objects read back byte for byte after the upgrade and after a rollback, with a clean stop and with kill -9. Existing objects keep their placement and aren't rebalanced.
• Crash, restart, corruption and eviction: no wrong bytes in any test. The connector branch's tests of these use the same striper.
• Soak: 30 minutes of reads at 21 GB/s while 2 GB/s of writes replaced the same objects. 323,029 loads were checked byte for byte, none wrong.
• Release checks: Rust, Python and end-to-end stages pass.