Skip to content
Krijn Doekemeijer edited this page Aug 23, 2023 · 1 revision

WAL designs

Open questions

Evaluation options

  • what is the micro-benchmark performance of read, write, append, and resets
  • what is the WAL performance on a single page writes (with flush!?) outside the application
    • what is the typical load generated in RocksDB for WAL
    • risk: limited by the hardware performance - all will behave the same with flush latencies
  • what the WAL replay time outside the application
  • 6.5.2 force tale latency experiment
  • 6.5.3 WA factor, is comparable so that is good.
  • 6.5.4 investigate reset calls? strange behavior
  • 6.5.5 heat-maps are cool but there is discrepancy - is it because of eager vs lazy zone reset management
  • concurrency performance
  • YCSB vs db_bench - one is only for RocksDB and other more general.

TODO

Animesh: give access to 2 node ZNS device and root access and compile 5.19. Reboot machine.

Krijn: NUMA setting clean. Check where interrupt ZNS device. Memory allocation NUMA. Check NUMA node of ZNS. Setup SPDK natively.

Clone this wiki locally