On 3/3/25 16:43, Colin Ian King wrote: > Modern processors can drop into deep sleep states relatively quickly > to save power. However, coming out of deep sleep states takes a small > amount of time and this is detrimental to performance for I/O devices > such as fast PCIe NVME drives when servicing a completed I/O > transactions. > > Testing with fio with read/write RAID0 PCIe NVME devices on various > modern SMP based systems (such as 96 thead Granite Rapids Xeon 6741P) > has shown that on 85-90% of read/write transactions issued on a CPU > are completed by the same CPU, so it makes some sense to prevent the > CPU from dropping into a deep sleep state to help reduce I/O handling > latency. For the platform you tested on that may be true, but even if we constrain ourselves to pci-nvme there's a variety of queue/irq mappings where this doesn't hold I'm afraid. > > This commit introduces a simple, lightweight and fast power sleep > demotion mechanism that provides the block layer a way to inform the > menu governor to prevent a CPU from going into a deep sleep when an > I/O operation is requested. While it is true that some I/Os may not s/requested/completed is the full truth, isn't it? > be serviced on the same CPU that issued the I/O request and hence > is not 100% perfect the mechanism does work well in the vast majority > of I/O operations and there is very small overhead with the sleep > demotion prevention. > > Test results on a 96 thread Xeon 6741P with a 6 way RAID0 PCIe NVME md > array using fio 3.35 performing random read and read-write test on a > 512GB file with 8 concurrent I/O jobs. Tested with the NHM_C1_AUTO_DEMOTE > bit set in MSR_PKG_CST_CONFIG_CONTROL set in the BIOS. > > Test case: random reads, results based on geometic mean of results from > 5 test runs: > Bandwidth IO-ops Latency Bandwidth > read (bytes/sec) per sec (ns) % Std.Deviation > Baseline: 21365755610 20377 390105 1.86% > Patched: 25950107558 24748 322905 0.16% What is the baseline? Do you mind trying with Rafael's recently posted series? Given the IOPS I'd expect good results from that alone already. https://lore.kernel.org/lkml/1916668.tdWV9SEqCh@xxxxxxxxxxxxx/ (Happy to see teo as comparison too, which you don't modify). > > Read rate improvement of ~21%. > > Test case: random read+writes, results based on geometic mean of results > from 5 test runs: > > Bandwidth IO-ops Latency Bandwidth > read (bytes/sec) per sec (ns) % Std.Deviation > Baseline: 9937848224 9477 550094 1.04% > Patched: 10502592508 10016 509315 1.85% > > Read rate improvement of ~5.7% > > Bandwidth IO-ops Latency Bandwidth > write (bytes/sec) per sec (ns) % Std.Deviation > Baseline: 9945197656 9484 288933 1.02% > Patched: 10517268400 10030 287026 1.85% > > Write rate improvement of ~5.7% > > For kernel builds, where all CPUs are fully loaded no perfomance > improvement or regressions were observed based on the results of > 5 kernel build test runs. > > By default, CPU power sleep demotion blocking is set to run > for 3 ms on I/O requests, but this can be modified using the > new sysfs interface: > > /sys/devices/system/cpu/cpuidle/psd_cpu_lat_timeout_ms rounding up a jiffie sure is a heavy price to pay then. > > setting this to zero will disabled the mechanism. > > Signed-off-by: Colin Ian King <colin.king@xxxxxxxxx> > --- > block/blk-mq.c | 2 + > drivers/cpuidle/Kconfig | 10 +++ > drivers/cpuidle/Makefile | 1 + > drivers/cpuidle/governors/menu.c | 4 + > drivers/cpuidle/psd.c | 123 +++++++++++++++++++++++++++++++ > include/linux/cpuidle_psd.h | 32 ++++++++ > 6 files changed, 172 insertions(+) > create mode 100644 drivers/cpuidle/psd.c > create mode 100644 include/linux/cpuidle_psd.h >
