Argon
Argon is a research storage server that combines cache partitioning, batched disk access and time quanta to limit interference between workloads.
Its R-value targets throughput relative to each workload's standalone rate within its allocated time share. Equal time shares do not imply equal byte rates. [2]
Cache partitions follow workload-specific profiles rather than equal-size division. Prefetching and write coalescing amortize positioning costs, while scheduling separates disk access. The design assumes CPU and network capacity are not bottlenecks. [2]
A cache-capacity shortfall can leave a newly added workload unsupported; the target is not unconditional admission. Longer transfers and quanta can increase worst-case response time. [1]
A required hit probability is not a byte allocation. Select a capacity from a workload-specific profile, then check the sum of exclusive partitions against available cache. The linked example distinguishes a budget shortfall from a profile with no qualifying sampled size; neither result is a modern-device admission guarantee. [3]
Relations
- related
- mClock, Slowdown fairness, Miss-ratio curve
Limitations
- A completion into write-back cache is not itself evidence of stable-media persistence. [1]