Intel engineer Francois Dugast today sent out the new patch series for enabling Transparent Hugepages (THP) support within the drm_pagemap code with a focus on the Intel Xe kernel driver usage. This enabling of THP support and in turn 2MB pages by the Xe driver is yielding "significant" performance improvements when using Shared Virtual Memory (SVM) such as for GPU compute workloads.

Dugast describes the patch series as providing "significant performance improvements when using SVM with 2MB pages."

He further elaborated on one of the patches within today’s patch series:

"This enables support for Transparent Huge Pages (THP) for device pages by using MIGRATE_VMA_SELECT_COMPOUND during migration. It removes the need to split folios and loop multiple times over all pages to perform required operations at page level. Instead, we rely on newly introduced support for higher orders in drm_pagemap and folio-level API.

In Xe, this drastically improves performance when using SVM. The GT stats below collected after a 2MB page fault show overall servicing is more than 7 times faster, and thanks to reduced CPU overhead the time spent on the actual copy goes from 23% without THP to 80% with THP:

Without THP:

svm_2M_pagefault_us: 966 svm_2M_migrate_us: 942 svm_2M_device_copy_us: 223 svm_2M_get_pages_us: 9 svm_2M_bind_us: 10

With THP:

svm_2M_pagefault_us: 132 svm_2M_migrate_us: 128 svm_2M_device_copy_us: 106 svm_2M_get_pages_us: 1 svm_2M_bind_us: 2"

Some nice work and complementing all of the other Intel Xe open-source driver improvements driven in part by Project Battlematrix as well as preparing for the upcoming Crescent Island AI inference accelerator and other hardware to come.

We’ll see if this v2 patch series in good shape for potentially landing in the upcoming Linux v6.20~7.0 cycle or if it will require more time to bake.