You cannot interpolate between the pure-read and pure-write numbers
The J750 is specified at 1,500 KIOPS 4KB random read and 300 KIOPS random write; the J770 doubles the write figure to 600 KIOPS. Weighting those by 70/30 produces a number that essentially never shows up on real hardware, because reads and writes do not share bandwidth — they share media resources. A program operation holds its target die for hundreds of microseconds, and every read landing on that die queues behind it. The real loss under a mixed load depends on how many dies the writes have turned into blocking points. Note up front that every J750/J770 figure quoted here is a product-definition/EVT planning value, used to convey magnitude and direction; the frozen release governs the final numbers.
Amplification starts at garbage collection
Thirty percent host writes is only the inbound figure. Once the drive reaches steady state, garbage collection must relocate valid data to free up blocks, so media-side writes equal host writes multiplied by write amplification. The pressure competing with reads for dies is therefore not 30 percent but 30 percent times WAF. Virtualisation and Ceph — target workloads for CORE 7 — are highly random with small fragmented blocks, where WAF typically runs well above what sequential writes produce.
Two usable levers
- Over-provisioning: usable capacity exchanged for a larger pool of free blocks. More free blocks mean the victim blocks garbage collection picks hold less valid data, and relocation volume falls accordingly. This is the most direct lever and the easiest to adjust at deployment time.
- Background pacing: advancing reclamation during troughs in host pressure and staying out of the way during bursts. It only works if the workload has troughs; a permanently saturated system has nothing to exploit.
Measurement has to follow the same mechanism
Because the loss originates in steady-state die contention, a mixed read/write score taken on a fresh drive carries no information. Precondition the full drive, wait until performance settles into its steady-state window before sampling, and record P99 and P99.9 write latency alongside the mean — a passing average with a diverging tail is precisely how die blocking presents itself under a mixed load.
What this means for selection
The J750 and J770 share a PCIe 4.0 x4 interface and the same 7,400 / 5,500 MB/s sequential figures. Their difference is concentrated in write capability: 1 DWPD against 3 DWPD, 300 KIOPS against 600. The higher the write fraction and the higher the amplification, the further apart the two drift. Selection should be driven by daily writes referred to the media side, not by the read/write ratio the host observes.