| Message ID | ZnMWDdKJHfYQLDzS@p100 (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Herbert Xu |
| Headers | show |
| Series | crypto: xor - fix template benchmarking | expand |
On 2024-06-19 1:31 p.m., Helge Deller wrote: > Commit c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking") > switched from using jiffies to ktime-based performance benchmarking. > > This works nicely on machines which have a fine-grained ktime() > clocksource as e.g. x86 machoines with TSC. > But other machines, e.g. my 4-way HP PARISC server, don't have such > fine-grained clocksources, which is why it seems that 800 xor loops > take zero seconds, which then calculates in the logs as: > > xor: measuring software checksum speed > 8regs : -1018167296 MB/sec > 8regs_prefetch : -1018167296 MB/sec > 32regs : -1018167296 MB/sec > 32regs_prefetch : -1018167296 MB/sec > > Fix this with some small modifications to the existing code to improve > the algorithm to always produce correct results without introducing > major delays for architectures with a fine-grained ktime() > clocksource: > a) Delay start of the timing until ktime() just advanced. On machines > with a fast ktime() this should be just one additional ktime() call. > b) Count the number of loops. Run at minimum 800 loops and finish > earliest when the ktime() counter has progressed. > > With that the throughput can now be calculated more accurately under all > conditions. > > Fixes: c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking") > Signed-off-by: Helge Deller <deller@gmx.de> You can add my "Tested-by". I wonder if prefetch versions are implemented correctly on parisc: [ 29.353868] xor: measuring software checksum speed [ 29.360030] 8regs : 2266 MB/sec [ 29.368031] 8regs_prefetch : 2076 MB/sec [ 29.376031] 32regs : 2259 MB/sec [ 29.384031] 32regs_prefetch : 2075 MB/sec [ 29.384080] xor: using function: 8regs (2266 MB/sec) > > diff --git a/crypto/xor.c b/crypto/xor.c > index 8e72e5d5db0d..29b4c0fd89d7 100644 > --- a/crypto/xor.c > +++ b/crypto/xor.c > @@ -83,33 +83,29 @@ static void __init > do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2) > { > int speed; > - int i, j; > - ktime_t min, start, diff; > + unsigned long reps; > + ktime_t min, start, t0; > > tmpl->next = template_list; > template_list = tmpl; > > preempt_disable(); > > - min = (ktime_t)S64_MAX; > - for (i = 0; i < 3; i++) { > - start = ktime_get(); > - for (j = 0; j < REPS; j++) { > - mb(); /* prevent loop optimization */ > - tmpl->do_2(BENCH_SIZE, b1, b2); > - mb(); > - } > - diff = ktime_sub(ktime_get(), start); > - if (diff < min) > - min = diff; > - } > + t0 = ktime_get(); > + /* delay start until time has advanced */ > + do { start = ktime_get(); } while (start == t0); > + reps = 0; > + do { > + mb(); /* prevent loop optimization */ > + tmpl->do_2(BENCH_SIZE, b1, b2); > + mb(); > + } while (reps++ < REPS || (t0 = ktime_get()) == start); > + min = ktime_sub(t0, start); > > preempt_enable(); > > // bytes/ns == GB/s, multiply by 1000 to get MB/s [not MiB/s] > - if (!min) > - min = 1; > - speed = (1000 * REPS * BENCH_SIZE) / (unsigned int)ktime_to_ns(min); > + speed = (1000 * reps * BENCH_SIZE) / (unsigned int)ktime_to_ns(min); > tmpl->speed = speed; > > pr_info(" %-16s: %5d MB/sec\n", tmpl->name, speed); >
On Wed, Jun 19, 2024 at 07:31:57PM +0200, Helge Deller wrote: > Commit c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking") > switched from using jiffies to ktime-based performance benchmarking. > > This works nicely on machines which have a fine-grained ktime() > clocksource as e.g. x86 machoines with TSC. > But other machines, e.g. my 4-way HP PARISC server, don't have such > fine-grained clocksources, which is why it seems that 800 xor loops > take zero seconds, which then calculates in the logs as: > > xor: measuring software checksum speed > 8regs : -1018167296 MB/sec > 8regs_prefetch : -1018167296 MB/sec > 32regs : -1018167296 MB/sec > 32regs_prefetch : -1018167296 MB/sec Ard, could you please take a look at this patch when you get a chance? Thanks!
On Wed, Jun 19, 2024 at 07:31:57PM +0200, Helge Deller wrote: > > + t0 = ktime_get(); > + /* delay start until time has advanced */ > + do { start = ktime_get(); } while (start == t0); Please rewrite this loop in the usual kernel coding style. What about adding a cpu_relax() in there if ktime_get doesn't advance? Something like while ((start = ktime_get()) == t0) cpu_relax(); Cheers,
On 7/6/24 02:04, Herbert Xu wrote: > On Wed, Jun 19, 2024 at 07:31:57PM +0200, Helge Deller wrote: >> >> + t0 = ktime_get(); >> + /* delay start until time has advanced */ >> + do { start = ktime_get(); } while (start == t0); > > Please rewrite this loop in the usual kernel coding style. Ok. > What about adding a cpu_relax() in there if ktime_get doesn't > advance? Something like > > while ((start = ktime_get()) == t0) > cpu_relax(); Yes, looks better. Will send updated patch. Helge
diff --git a/crypto/xor.c b/crypto/xor.c index 8e72e5d5db0d..29b4c0fd89d7 100644 --- a/crypto/xor.c +++ b/crypto/xor.c @@ -83,33 +83,29 @@ static void __init do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2) { int speed; - int i, j; - ktime_t min, start, diff; + unsigned long reps; + ktime_t min, start, t0; tmpl->next = template_list; template_list = tmpl; preempt_disable(); - min = (ktime_t)S64_MAX; - for (i = 0; i < 3; i++) { - start = ktime_get(); - for (j = 0; j < REPS; j++) { - mb(); /* prevent loop optimization */ - tmpl->do_2(BENCH_SIZE, b1, b2); - mb(); - } - diff = ktime_sub(ktime_get(), start); - if (diff < min) - min = diff; - } + t0 = ktime_get(); + /* delay start until time has advanced */ + do { start = ktime_get(); } while (start == t0); + reps = 0; + do { + mb(); /* prevent loop optimization */ + tmpl->do_2(BENCH_SIZE, b1, b2); + mb(); + } while (reps++ < REPS || (t0 = ktime_get()) == start); + min = ktime_sub(t0, start); preempt_enable(); // bytes/ns == GB/s, multiply by 1000 to get MB/s [not MiB/s] - if (!min) - min = 1; - speed = (1000 * REPS * BENCH_SIZE) / (unsigned int)ktime_to_ns(min); + speed = (1000 * reps * BENCH_SIZE) / (unsigned int)ktime_to_ns(min); tmpl->speed = speed; pr_info(" %-16s: %5d MB/sec\n", tmpl->name, speed);
Commit c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking") switched from using jiffies to ktime-based performance benchmarking. This works nicely on machines which have a fine-grained ktime() clocksource as e.g. x86 machoines with TSC. But other machines, e.g. my 4-way HP PARISC server, don't have such fine-grained clocksources, which is why it seems that 800 xor loops take zero seconds, which then calculates in the logs as: xor: measuring software checksum speed 8regs : -1018167296 MB/sec 8regs_prefetch : -1018167296 MB/sec 32regs : -1018167296 MB/sec 32regs_prefetch : -1018167296 MB/sec Fix this with some small modifications to the existing code to improve the algorithm to always produce correct results without introducing major delays for architectures with a fine-grained ktime() clocksource: a) Delay start of the timing until ktime() just advanced. On machines with a fast ktime() this should be just one additional ktime() call. b) Count the number of loops. Run at minimum 800 loops and finish earliest when the ktime() counter has progressed. With that the throughput can now be calculated more accurately under all conditions. Fixes: c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking") Signed-off-by: Helge Deller <deller@gmx.de>