Understanding Storage IO Aggregation, RAID Write Penalties, and Performance Evaluation Models

The previous article introduced basic performance evaluation concepts; this follow‑up focuses on practical configuration guidance, highlighting the challenges of end‑to‑end analysis from host ports to storage subsystems and presenting common pitfalls in performance testing.

IO aggregation and full‑stripe writes : When IO is aggregated to a full‑stripe size, pre‑read is unnecessary and RAID write penalties are avoided. For RAID5‑5, a single data write triggers two pre‑reads and one parity write, effectively expanding one IO into four. Full‑stripe writes combine four data IOs into five, greatly improving efficiency.

Factors influencing IO merge capability : (1) Host‑side IO model – sequentiality and contiguity depend on the host OS, block device, volume manager, and HBA configuration. (2) Storage‑side merge ability – caches, disk modules, and controllers attempt to coalesce small IOs into larger ones before writing to disk.

Typical workload models :

OLTP – small random IO (≈8 KB), read/write ratio ~3:2, latency 10‑20 ms.

OLAP – large sequential IO (≈512 KB), >90 % reads, mixed read/write on temporary LUNs.

VDI – mixed read/write, latency ~10 ms, small IO ranging 512 B‑16 KB.

SPC‑1 – industry‑standard random IO benchmark, read/write ratio ~4:6, IO size ≈4 KB.

SSD, SAS, NL‑SAS comparison : The article briefly mentions performance characteristics of different drive types.

FC link bandwidth calculation : Provides formulas for 8 Gbps FC, accounting for 8b/10b encoding, class‑3 efficiency (97.15 %), and shows how to compute theoretical MB/s and real‑world limits.

Impact of parity and write penalties : In RAID5‑5 (4D+1P) each four data IOs incur one parity IO; in RAID6‑6 (4D+2P) the penalty is higher. Write penalties reduce effective bandwidth, especially for random write‑heavy workloads.

Read/write ratio importance : Higher write ratios increase resource consumption due to longer IO paths, cache mirroring, and parity updates. The article quantifies the effect using examples.

RAID level performance and capacity trade‑offs : Shows that RAID10, RAID5, and RAID6 deliver different effective performance and usable capacity, with RAID6 offering the best reliability but lower performance for write‑intensive workloads.

Sequential vs. random IO : Sequential IO benefits from lower seek times and higher throughput; random IO suffers from mechanical latency but can be mitigated by caching and IO merging.

IO size impact : Small IO (< 16 KB) is measured in IOPS, large IO (≥ 32 KB) in bandwidth. SPC‑1 focuses on small random IO, SPC‑2 on large sequential workloads.

Cache acceleration : Describes write‑back caching, write‑hit optimization, and write‑merge, as well as read‑hit benefits and full‑cache hits that eliminate disk access entirely.