How IO Aggregation, RAID Choices, and Workload Types Shape Storage Performance

1. IO Aggregation and Write Penalty

When IO is aggregated to full‑stripe size, pre‑read is unnecessary and RAID write penalty is avoided; otherwise, RAID5‑5 small‑write triggers two pre‑reads and one parity write, effectively expanding one logical IO into four physical IOs. Full‑stripe writes combine four data IOs plus one parity write, improving efficiency.

Storage’s ability to merge IO depends on two factors:

Host‑side IO order and continuity, influenced by host software, block devices, volume management, and HBA policies.

Cache, block devices, and disks along the IO path that attempt to sort and merge small IOs into larger ones.

Random database workloads often cannot achieve full‑stripe merging, leading to varying merge efficiency.

2. Business Models and Their Characteristics

The article distinguishes four common workload models used for performance evaluation:

OLTP : Small transactional reads/writes (≈8 KB), high concurrency, latency 10‑20 ms, read/write ratio ~3:2.

OLAP : Large, complex queries lasting minutes to days, predominantly read‑heavy (>90 % reads), mixed large sequential IO.

VDI : Startup, login, and steady‑state phases; startup is read‑intensive, login is write‑intensive, steady state has ~2:8 read/write ratio with IO sizes 512 B‑16 KB.

SPC‑1 : Industry‑standard random IOPS benchmark; mixed read/write (≈4:6), 4 KB random IO, with distinct ASU regions (data, user, log).

3. Impact of Checksum (Parity) on Performance

In RAID5‑5 (4D+1P) each four data writes generate one parity write, consuming bandwidth without benefiting the host workload. Effective write bandwidth is calculated as:

Effective Disk Write BW = Single‑Disk Sequential Write BW × Disk Count × (Data Disks / Total Disks)
Product Effective Write BW = MIN(Product Max Write BW, Effective Disk Write BW)

Example: 96 × 600 GB 15K SAS disks (30 MB/s each) in RAID6‑6 (4D+2P) yields 1920 MB/s effective write bandwidth.

4. Read/Write Ratio Influence

Higher write ratios increase resource consumption and reduce effective performance. Write‑heavy workloads also incur longer IO paths (mirroring, write‑hole protection) and higher latency.

5. RAID Level Performance Differences

RAID10, RAID5, and RAID6 are compared under identical disk counts and workload ratios. Write‑heavy scenarios suffer more on RAID5/6 due to parity overhead. Capacity‑wise, RAID6 offers the best reliability, RAID10 the next, and RAID5 the least.

6. Sequential vs. Random IO Characteristics

Sequential IO outperforms random IO on both mechanical disks and SSDs. Small sequential IO can be merged into larger IOs, achieving higher IOPS, while random small IO suffers low cache hit rates and cannot be merged efficiently.

7. IO Size Effects

Small IOs are measured by IOPS, large IOs by bandwidth. Increasing random IO size beyond 16 KB on HDDs reduces IOPS linearly. Storage systems often coalesce multiple small IOs into a larger one (e.g., sixteen 8 KB writes become one 128 KB write) to improve throughput.

8. Cache (Cache) Impact

Cache accelerates both read and write paths:

Write‑back: Data is first written to cache, allowing asynchronous bulk flushes to disk, roughly doubling write performance.

Write hit: Subsequent writes to the same address hit cached data, reducing disk writes.

Write merge: Multiple small writes are combined into a larger write before reaching the disk.

Read cache hit: Frequently accessed data served entirely from cache (full hit) yields the shortest latency and maximum IOPS.

Understanding these factors enables accurate performance assessment and optimal storage configuration for specific business workloads.