Work Continues For Supporting Intel Sapphire Rapids C0.x Idle States On Linux
While Intel's Linux engineers were very timely in enabling much of the Xeon Scalable "Sapphire Rapids" support for the upstream Linux kernel well ahead of the processor launch at the beginning of the year, one patch series that has continued on post-launch has been working to get the new C0.x idle states supported.
Sapphire Rapids introduces new C0.1/C0.2 idle states that are between the most shallow POLL idle state and C1 as traditionally the next lower power state. The C0.1 and C0.2 idle states offer a mix of advantages between POLL and C1 by having lower wake-up latency while saving more power than POLL. These C0.x idle states were engineered for latency-sensitive workloads.
Posted today to the Linux kernel mailing list is the fourth iteration of the patches enabling the C0.x idle states. Various code review issues were addressed this round.
The focus on these patches is for making use of the C0.2 idle state and would be enabled by default for Sapphire Rapids and newer while could be disabled via the intel_idle.states_off=2 module parameter. From POLL to C0.2 a 13% drop in AC power was noted or 18% in the CPU RAPL power consumption. With Hackbench tests the shallower idle state allowed for a 0.6~4% improvement.
These patches are too late for the v6.5 kernel but perhaps we'll finally see this support land later in the year with v6.6. Those wanting these C0.x idle state patches for Sapphire Rapids right now can find them on the kernel mailing list.
Sapphire Rapids introduces new C0.1/C0.2 idle states that are between the most shallow POLL idle state and C1 as traditionally the next lower power state. The C0.1 and C0.2 idle states offer a mix of advantages between POLL and C1 by having lower wake-up latency while saving more power than POLL. These C0.x idle states were engineered for latency-sensitive workloads.
Posted today to the Linux kernel mailing list is the fourth iteration of the patches enabling the C0.x idle states. Various code review issues were addressed this round.
The focus on these patches is for making use of the C0.2 idle state and would be enabled by default for Sapphire Rapids and newer while could be disabled via the intel_idle.states_off=2 module parameter. From POLL to C0.2 a 13% drop in AC power was noted or 18% in the CPU RAPL power consumption. With Hackbench tests the shallower idle state allowed for a 0.6~4% improvement.
These patches are too late for the v6.5 kernel but perhaps we'll finally see this support land later in the year with v6.6. Those wanting these C0.x idle state patches for Sapphire Rapids right now can find them on the kernel mailing list.
Add A Comment