Some stats from my tweaking gossip intervals to understand their impact, seeing how long it takes to upload files of various sizes.
| Baby-fleming | Change A | Change B | |
|---|---|---|---|
| Size (kB) | Time (s) | Time (s) | Time (s) |
| 700 | 11.271 | 49.47 | 21.688 |
| 800 | 12.074 | 53.454 | 16.894 |
| 900 | 14.047 | 60.628 | 19.654 |
| 1000 | 14.595 | 57.409 | 19.647 |
| 1100 | 13.292 | 50.306 | 23.107 |
| 1200 | 85.711 | 62.416 | 19.867 |
| 1300 | 23.079 | 67.187 | 24.865 |
| 1400 | 32.758 | 48.352 | 22.146 |
| 1500 | killed for cpu | 74.33 | 26.145 |
| 1600 | 49.884 | 25.31 | |
| 1700 | 46.471 | 27.088 | |
| 1800 | 68.727 | 24.599 | |
| 1900 | 50.433 | 29.201 | |
| 2000 | 46.137 | 30.306 | |
| 2200 | 60.206 | 26.344 | |
| 2400 | 70.675 | 42.646 | |
| 2600 | 57.213 | 62.955 | |
| 2800 | 74.288 | 49.274 | |
| 3000 | 360.143 | 93.373 | |
| 3500 | 63.956 | 66.507 | |
| 4000 | 81.542 | 57.404 | |
| 4500 | 101.312 | killed for cpu | |
| 5000 | 65.263 | ||
| 6000 | 81.199 | ||
| 7000 | 182.018 | ||
| 8000 | 162.767 | ||
| 9000 | 229.971 | ||
| 10000 | killed for ram |
Change A: Gossip slower, instead of 1s updates change to 7s updates
src/parsec.rs:L63 (or use ROUTING_GOSSIP_PERIOD environment variable)
-pub const GOSSIP_PERIOD: Duration = Duration::from_secs(1);
+pub const GOSSIP_PERIOD: Duration = Duration::from_secs(7);
Allowed larger files to complete but was overall slower for all file sizes. Shows that maybe nodes are being flooded by other gossip so can’t make up their mind about their own stuff…?
Change B: Gossip spread out between 1-3s rather than always every 1s, to try to help any lockstep situations that may be happening.
src/parsec.rs:L364
pub fn gossip_period(&self) -> Duration {
- self.gossip_period
+ let t = rand::thread_rng().gen_range(0, 2000);
+ self.gossip_period + Duration::from_millis(t)
}
Allowed larger files to be uploaded without the performance hit from change a, but files were not as large as change a.