Three phases, one day
A virtual desktop workload is not a flat curve. Split by time of day it resolves into three phases whose demands scarcely overlap:
- boot storm — hundreds of desktops powering on at the morning peak, dominated by 4K–32K random reads at very high instantaneous queue depth;
- steady-state office hours — mixed read/write with a rising write share, small and scattered blocks;
- maintenance window — patch distribution, full antivirus scans, image refreshes: bulk homogeneous writes, and the highest write amplification of the day.
The storm is a tail-latency test
Boot storms get mistaken for a bandwidth problem. What they actually test is 4K random read capability at depth, and the shape of its distribution. The J750/J770 target 1,500 KIOPS on 4K random read with 65 µs average QD1 read latency — but what users perceive is P99.9. A single login chains dozens of small reads; if any one of them stalls for hundreds of milliseconds, the user experiences "logins are slow today". Drives that meet the mean but lose control of the tail perform worst here.
Where steady-state writes come from
Writes come mostly from linked-clone delta disks, roaming user profiles, page files and temp directories. They are small, random, and repeatedly overwrite the same logical regions — unfriendly to garbage collection, so steady-state write amplification runs well above sequential cases. Endurance selection therefore follows desktop type: for largely read-only images with small deltas, the J750 at a planned 1 DWPD suffices; for persistent desktops keeping user data local, the J770 at a planned 3 DWPD and 600 KIOPS random write fits better.
Form factor and density
Desktop density sets the per-node capacity point. U.2/U.3 15mm suits mainstream 2U chassis and remains the default for most VDI nodes. For high-density nodes or OCP racks, FLEX 9 in E3.S (J100E/J130E) and E1.S (J100S/J130S) fits more drives into the same volume, at the cost of a tighter per-drive power budget and thermal design that must be settled early.
Two things that get missed
First, power-loss protection. An unplanned outage on a VDI node hits hundreds or thousands of sessions at once, and unflushed write-buffer contents must be covered by PLP — baseline, not an option. Second, namespace partitioning: putting the golden image and the delta disks in separate namespaces keeps the mapping and reclaim behaviour of two very different access patterns from interfering. These capabilities are being validated on the CORE 7 platform at product-definition/EVT stage.