Re: [PATCH] powerpc/defconfigs: Set HZ=1000 on ppc64 and powernv defconfigs

[Srikar Dronamraju <srikar@linux.ibm.com>]

* Madadi Vineeth Reddy <vineethr@linux.ibm.com> [2025-03-30 13:17:34]:

> Commit 030bdc3fd080 ("powerpc/defconfigs: Set HZ=100 on pseries and ppc64
> defconfigs") lowered CONFIG_HZ from 250 to 100, citing reduced need for a
> higher tick rate due to high-resolution timers and concerns about timer
> interrupt overhead and cascading effects in the timer wheel.
> 
> However, improvements have been made to the timer wheel algorithm since
> then, particularly in eliminating cascading effects at the cost of minor
> timekeeping inaccuracies. More details are available here
> https://lwn.net/Articles/646950/. This removes the original concern about
> cascading, and the reliance on high-resolution timers is not applicable
> to the scheduler, which still depends on periodic ticks set by CONFIG_HZ.
> 
> With the introduction of the EEVDF scheduler, users can specify custom
> slices for workloads. The default base_slice is 3ms, but with CONFIG_HZ=100
> (10ms tick interval), base_slice is ineffective. Workloads like stress-ng
> that do not voluntarily yield the CPU run for ~10ms before switching out.
> Additionally, setting a custom slice below 3ms (e.g., 2ms) should lower
> task latency, but this effect is lost due to the coarse 10ms tick.
> 
> By increasing CONFIG_HZ to 1000 (1ms tick), base_slice is properly honored,
> and user-defined slices work as expected. Benchmark results support this
> change:
> 
> Latency improvements in schbench with EEVDF under stress-ng-induced noise:
> 
> Scheduler       CONFIG_HZ  Custom Slice  99th Percentile Latency (µs)
> --------------------------------------------------------------------
> EEVDF           1000       No            0.30x
> EEVDF           1000       2 ms          0.29x
> EEVDF (default) 100        No            1.00x
> 
> Switching to HZ=1000 reduces the 99th percentile latency in schbench by
> ~70%. This improvement occurs because, with HZ=1000, stress-ng tasks run
> for ~3ms before yielding, compared to ~10ms with HZ=100. As a result,
> schbench gets CPU time sooner, reducing its latency.
> 
> Daytrader Performance:
> 
> Daytrader results show minor variation within standard deviation,
> indicating no significant regression.
> 
> Workload (Users/Instances)  Throughput 1000HZ vs 100HZ (Std Dev%)
> --------------------------------------------------------------------------
> 30 u, 1 i                   +3.01% (1.62%)
> 60 u, 1 i                   +1.46% (2.69%)
> 90 u, 1 i                   –1.33% (3.09%)
> 30 u, 2 i                   -1.20% (1.71%)
> 30 u, 3 i                   –0.07% (1.33%)
> 
> Avg. Response Time: No Change (=)
> 
> pgbench select queries:
> 
> Metric                         1000HZ vs 100HZ (Std Dev%)
> ------------------------------------------------------------------
> Average TPS Change             +2.16% (1.27%)
> Average Latency Change         –2.21% (1.21%)
> 
> Average TPS: Higher the better
> Average Latency: Lower the better
> 
> pgbench shows both throughput and latency improvements beyond standard
> deviation.
> 
> Given these results and the improvements in timer wheel implementation,
> increasing CONFIG_HZ to 1000 ensures that powerpc benefits from EEVDF’s
> base_slice and allows fine-tuned scheduling for latency-sensitive
> workloads.
> 
> Signed-off-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com>

Good work Vineeth,

As pointed by you, the base slice is 3ms and having base slice as a multiple
of tick will help.  The numbers also support this change.

Looks good to me.

Reviewed-by: Srikar Dronamraju <srikar@linux.ibm.com>

-- 
Thanks and Regards
Srikar Dronamraju