Low disk performance improvement with Samsung RAPID?

[Willshak]

Hi,

Samsung SSDs can use a "RAPID" mode where host memory is used as an SSD cache. For some workloads this can provide a massive performance benefit to the physical machine. It also helps disk performance inside the virtual machine, but to a much smaller degree:

PassMark Disk Benchmarks (MB/s)

______________________________  Random RW   Seq Read   Seq Write
Physical, No RAPID               200         277        231
Physical, RAPID                 2072        2526       2123
Guest, No RAPID, No Host Cache   193         289        203
Guest, RAPID, No Host Cache      276         330        350
Guest, RAPID, Host Cache         326         315        334

So the physical host gets a whopping 10x speed-up in these (synthetic, your mileage will vary etc.) benchmarks, but the VirtualBox guest only gets about 40% speed-up.

Any idea if this is as expected, or if more of the "RAPID" caching effect can be made available inside the guest?

Thanks,
Kristian

• Host: Windows 7 SP1 64-bit desktop with i7 4770K, 3.5GHz, 4 physical cores, HT enabled, 16GB memory
  Host Disk: Samsung 840 EVO 1TB SSD on SATA3
• Guest: Windows Server 2008 R2 Standard SP1 64-bit, 8GB memory, both 4 and 8 processors tested
  Guest Storage: .VDI dynamic storage file (pre-grown before benchmarks) on SATA Controller
  Guest Settings: IO-APIC, 100% Execution Cap, PAE/NX, VT-X / AMD-V, Nested Paging all enabled

[Willshak]

One addition: using a shared folder is (in my particular set-up) some 19% faster than accessing the .VDI. Still a long way away from the physical host though.

_______________________________   Random RW   Seq Read   Seq Write
Guest, RAPID, Host Cache, .VDI     326         315        334
Guest, RAPID, Host Cache, Share    300         430        430      <<< Added

[mpack]

I/O benchmarks run inside a guest are not reliable. The timer is shared with the host, and most of the actual I/O will take place in a host process.

[Willshak]

Each benchmark might run for 15 seconds, so hopefully timer accuracy shouldn't be an issue, and the order of magnitude difference in caching effect seems quite stable across multiple runs as well as across multiple different benchmark tools.

That said, these are still synthetic benchmarks, and obviously only actual workload performance really matters...