This is a read-only mirror of the HSC engineering wiki, restored from a 2017 archive. Some links are broken and some content is out of date. About this mirror.
Cluster Sizing Worksheet
From ANTFARM Wiki
| Cluster Sizing Worksheet | |
|---|---|
| Purpose | estimate forager count for a workload |
| Inputs | tasks/hour, model share, split |
| Status | current |
| Archive ref | QUANTARA-SWARMGLASS-R11-17EA3F |
Contents[hide] |
A back-of-envelope worksheet for deciding how many foragers a workload needs. Derived from Distributed Inference Notes/Appendix B numbers; do not use it for anything that matters without re-measuring.
[edit] Inputs
| Symbol | Meaning | Typical |
|---|---|---|
| T | tasks per hour | 2,000 |
| m | fraction of tasks that call a model | 0.15 |
| t_c | mean cpu/io task seconds | 4 |
| t_m | mean model task seconds (Basalt-2, split=1) | 9 |
| s | slots per forager | 2 |
| u | target utilisation | 0.6 |
[edit] Formula
busy_seconds_per_hour = T * ((1 - m) * t_c + m * t_m) foragers_needed = busy_seconds_per_hour / (3600 * s * u)
With the typical column: 2000 × (0.85 × 4 + 0.15 × 9) = 9,500 s/h → 9,500 / (3600 × 2 × 0.6) ≈ 2.2 foragers. The colony ran eleven. The extra capacity was for the index rebuilds, not steady state.
[edit]
| m | foragers |
|---|---|
| 0.05 | 2.0 |
| 0.15 | 2.2 |
| 0.40 | 2.8 |
| 0.80 | 3.7 |
[edit] Caveats
- Ignores warmup, which dominates after any shard restart.
- Ignores the bus, which dominates at split ≥ 2.
- Assumes
assign_timeoutnever fires. It fires.
— svanterpool 2015-11-05
