HN Explorer

7 comments

[ 4.0 ms ] story [ 25.0 ms ] thread
apankrat
Just a very pedantic IO test that goes through a long list of permutations of IO buffer sizes, counts and IO modes and finds the one that delivers the highest read/write rate. Basically, it estimates maximum IO rates realistically achievable by a userspace process, as opposed to a physical read/write rate at a disk level, which can be (misleadingly) higher.

The theory is over at https://ccsiobench.com/fast-bulk-io

    ---
NVMe drive from my test box - https://ccsiobench.com/s/UAr3f

SSD in USB3 enclosure - https://ccsiobench.com/s/UAtNj

TrueCrypt volume on the same - https://ccsiobench.com/s/UAxmh

Sunset
Will this work with something like Ext2Fsd or other usermode file system drivers not from microsoft which work on windows?
apankrat
Yep, it works with anything that is addressable as a file system location, i.e. as X:\Path or \\Server\Share\Path format.
kazinator
Having measured file system performance on Windows, what action can you take based on the results?
apankrat
For one, you can get an idea if you are getting proper performance out of your devices. For example, NVMe drives plugged directly into certain revisions of ASYS motherboards will not get to their spec speeds, you will need to use a M2/PCIe raiser card. Similarly, not all USB3 ports on motherboards are created equal. I have about 20% difference in performance of external drive between two ports on my work machine.

For two, you can estimate the overhead you get by enabling NTFS compression and/or Bitlocker for a drive.

Ditto for the antivirus overhead. For example, BitDefender is known to absolutely kill multi-stream SMB3 performance, dragging 10GbE speeds down to 1Gbps or less.

ferongr
I think the parent meant, what can you tune in Windows itself to get the "best" performance shown in the tool? Or are those buffer sizes and counts program dependent and not something the user can change?
apankrat
Short of enabling jumbo frames on network adapters I don't think there's much in terms of system-wide tuning. Every program that cares about IO performance would implement it on top of raw File API (CreateFile, ReadFile, etc), so the IO strategy will indeed be concealed in the guts of the program.