Twenty Four Spindles

You are laying out 24 identical disks for a metrics store. Measured workload:

  • 95 percent of the bytes arrive as sequential appends of 8 MB segments, written whole
  • 5 percent is small random reads serving dashboard queries
  • the array must keep serving through one disk failure
  • usable capacity per dollar is a stated goal

Which layout fits best?

A

RAID-0. Segments are already replicated at the application layer by the metrics store, so the array should be tuned purely for the N⋅SN \cdot S sequential bandwidth.

B

RAID-5. The 8 MB appends land as full-stripe writes, so parity is computed from the incoming data with no read-modify-write, sequential bandwidth is (N−1)⋅S(N-1) \cdot S, usable capacity is 23 of 24 disks, and rotated parity lets random reads use all 24 spindles.

C

RAID-4. The workload is almost entirely full-stripe writes, so the dedicated parity disk is never a bottleneck, and a fixed parity disk is simpler to build and to rebuild.

D

RAID-1 across 12 mirrored pairs. Mirroring has no parity arithmetic on the write path at all, which matters most when 95 percent of the work is writing.