ZFS On Linux 0.7.13 Released With Fixes For Linux 5.0 Kernel Compatibility

[phoronix]

Phoronix: ZFS On Linux 0.7.13 Released With Fixes For Linux 5.0 Kernel Compatibility

While we are very much looking forward to the huge ZFS On Linux 0.8 release, as a new stable release for offering up compatibility with the newly minted Linux 5.0 is now the ZoL 0.7.13 milestone...

http://www.phoronix.com/scan.php?page=news_item&px=ZFS-On-Linux-0.7.13-Released

[Chugworth]

I'd like to see some benchmarks on how the ZoL 0.7.13 release runs on the 5.0 kernel compared with 4.20.

[Michael]

I'd like to see some benchmarks on how the ZoL 0.7.13 release runs on the 5.0 kernel compared with 4.20.

AFAIK, no real performance changes are expected for 4.20 vs. 5.0?

[Chugworth]

AFAIK, no real performance changes are expected for 4.20 vs. 5.0?

Well I'm no expert in kernel programming, but the fix that the ZoL developers described makes it sound to me like there could be a significant performance hit:

The 5.0 kernel no longer exports the functions we need to vector
(SSE/SSE2/SSE3/AVX...) instructions. Disable vector-based checksum
algorithms when building against those kernels.

It would also be interesting to throw FreeBSD 12 into the ZFS benchmarking mix.

[Yoshi]

But shouldn't it get benchmarked against 4.19? I thought 4.20 already contains these "fixes"...

[darkbasic]

But shouldn't it get benchmarked against 4.19? I thought 4.20 already contains these "fixes"...

Exactly. Michael can you please run some benchmarks?

[Chugworth]

But shouldn't it get benchmarked against 4.19? I thought 4.20 already contains these "fixes"...

The stuff I've read about the issue mentions 5.0 as the affected kernel. I didn't realize the changes were already in 4.20, but I do see the previous ZoL release only supported through 4.19.

[mercutio]

the ZoL code checks if the new broken behaviour is there or not rather than Linux version. So if you fix your linux kernel source then the performance regression should be able to be bypassed. That said, it's in the checksumming code that the CPU optimisations are used predominantly afaik, which already run damn fast, so it's probably cpu usage %, power usage that goes up.

[Chugworth]

the ZoL code checks if the new broken behaviour is there or not rather than Linux version. So if you fix your linux kernel source then the performance regression should be able to be bypassed. That said, it's in the checksumming code that the CPU optimisations are used predominantly afaik, which already run damn fast, so it's probably cpu usage %, power usage that goes up.

Yeah, I wondered if the benchmarks would look the same but the CPU usage goes up. I will also be curious to know if Ubuntu would modify their kernel to allow ZoL that access.