Memory is slow, Disk is fast
Web page · https://bitflux.ai/blog/memory-is-slow-part3
First seen by Oriel on .
AI summary
The article compares an in-RAM AVL tree with DiskArray, finding that a batched, disk-based sorted structure outperforms the tree beyond 16M keys and scales to 64G nodes.
- DiskArray stores sorted keys in 4 KiB extents, keeps a compact extent index in RAM, and limits inserts and deletes to shifting keys within an extent.
- It queues 2048 transactions and uses 28 io_uring workers across eight NVMe drives to overlap reads, updates, and writes.
- At 2G nodes, DiskArray reaches 670K inserts/s versus 320K for AVL, despite moving more data per insert.
- AVL requires 72 bytes of RAM per key and 144 GiB at 2G nodes; DiskArray uses disk storage and a small buffer pool.
- The author notes a top-level index corner case that can cause occasional O(n) work, and argues Big O alone misses hardware latency and bandwidth effects.
#diskarray #nvme #data structures #io_uring
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Lobsters: 10, latest observation ; +9 over observed period since
Comments
Comment count reported by the source.
- Lobsters: 8, latest observation ; +7 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Lobsters | pvfzhq | https://www.bitflux.ai/blog/memory-is-slow-part3/ Read discussion |