Research on temporal determinism (stability) of current OS bypass methods.
We want to evaluate latency and jitter in datacenter communication, using different kernel bypass methods. Moreover, we want to determine the system factors impacting jitter (determinism/stabililty). We focus particularly on periodic communication.
Tools:
- no bypass
- RDMA
- DPDK
- eBPF XDP (to app: map polling, socket)
Timeframes:
- packet numbers: 10k,1mil, more
- absolute time: up to days
Benchmark variables:
- Endhost settings: {no isol, isol}, {intervals}, {interrupts shielding} {memory /paging}
- network settings: p2p, scale up
Evaluation methods:
- single/round trip latency: latency_buckets, update_max, packet_dump
- metrics: max, avg, 99.x percentile
- jitter locality: TX stack, RX stack, network
- one way jitter: send and recv difference in periodic protocols