Removal of an obsolete AMD fix makes Linux kernel 6 faster
Since 2002, Linux performance has been suffering as a result of an outdated workaround for power management issues on AMD computers. It's gone now.
Modern silicon processors have the advantage that power control is crucial. Since the 20th century, power conservation and battery life have not been a concern. The only reason that processor manufacturers can continue to make money is by selling us more and more transistors because, despite excellent cooling, most of them are off most of the time.
Software OSes
Linux kernel 6 runs faster after removing an outdated AMD patch.
After 20 years, a performance-degrading workaround was rediscovered.
Since 2002, Linux performance has been suffering as a result of an outdated workaround for power management issues on AMD computers. It's gone now.
Modern silicon processors have the advantage that power control is crucial. Since the 20th century, power conservation and battery life have not been a concern. Because most transistors in processors are switched off the majority of the time, no matter how well cooled they are, processor vendors must continually sell us more transistors to keep their business afloat.
Around that time, an issue was discovered with some AMD Athlon systems using VIA chipsets, a year after the Linux kernel received ACPI support. The problematic computers waited a while for the STPCLK# signal to switch a CPU core to idle (albeit there was only one core back then), so the kernel developers inserted some dummy I/O read instructions to make sure the processor wouldn't keep running when it was supposed to be shutting. It enhanced power management and compatibility.
The issue is that Linux still executes it on AMD processors, as AMD engineer K Prateek Nayak discovered lately. He discovered the problem when examining the behavior of the kernel using AMD's "Instruction-Based
As a result, Intel's Dave Hansen has rushed out a fix that leaves the workaround in place but only applies to Intel processors, where it shouldn't be a problem because Intel chips use a different mechanism to transition to idle.
Hansen clarifies:
On current platforms, this workaround is rather painful. It may take millions of cycles to run "inl()".
The lowest throughput increases by around 14 times, and the mean throughput by little over half, according to the patch's benchmark statistics.
The patch has already been approved and should be applied, even though kernel 6.0 is scheduled to release on Sunday. It will be fascinating to see how much of a difference it makes in performance in the actual world.
In an era where there are mitigations for current CPU issues
No comments: