Benchmark: Raft behavior while testing Membership() throughput

BBVA / qed

The scalable, auditable and high-performance tamper-evident log project

https://qed.readthedocs.io/

Apache License 2.0

95 stars 19 forks source link

Benchmark: Raft behavior while testing Membership() throughput #53

Closed jllucas closed 5 years ago

jllucas commented 5 years ago

We need to measure the system write/query performance when using raft replication, and querying only 1 follower (instead of both).

We should consider the following scenarios: Multi node:

[x] One Leader & two followers{1..2}
- Start the cluster
- Preload N events to the leader
- Query Membership of the N events to follower1
- Perform continuous write load on the leader
- At the same time, perform Query Membership of the previous N events ONLY to follower 1 (notice that follower 2 is query-load free)

suizman commented 5 years ago

Here you can find the results for this test case:

We've started 1 leader 2 followers
Preload 100K events
Query Membership only to follower1 without additional load
While performing continuous write load to the leader perform reads only on follower1

We can observer that the write performance goes down form ~2150/s to ~1950/s while maintaining around 1250/s of reads on the follower1 and follower2 remains with 0% load.

See attachment below: MULTINODE2-READ-FOLLOWER1.txt