Linux 7.4 to Speed Up Software RAID: Optimizing xor_gen for AVX-512

The upcoming Linux 7.4 development cycle is set to bring in an AVX-512-optimized implementation of the xor_gen() function. This function is used by the software RAID code and is responsible for generating and verifying parity blocks — in particular, for RAID5 and RAID6 arrays. According to Phoronix, the new version delivers a tangible performance boost.

The patches for this work were in the making for several months, and they have now finally been readied for submission. The effort is led by Eric Biggers of Google — an engineer who in recent years has already repeatedly sped up the Linux kernel's cryptographic code by drawing on the capabilities of AVX-512 and other modern instruction sets.

The gains will benefit server and desktop processors with AVX-512 support: AMD chips on the Zen 4 architecture and newer, as well as current Intel server CPUs. As an example of the improvement, the article cites a test platform based on the AMD Ryzen 9 9950X.

After another round of revision, the patches landed in the x86/fpu branch of the TIP repository and have been queued for merging. The merge window for Linux 7.4 will open in the second half of October, so barring any last-minute surprises, the optimization will make it into exactly this kernel release.

For users of software RAID arrays, the change is a practical one: the faster parity blocks are computed and checked, the lighter the load on the processor when working with disk arrays. And no special action will be required — the kernel itself will select the accelerated implementation on suitable hardware.

Source: Phoronix