Skip to contentSkip to conversation

crypto: xor - fix template benchmarking

PATCH In review Helge Deller 2 participants 2 messages Jul 8, 2024 – Aug 2, 2024 Crypto API

lore
Share & tools
lore

Helge Deller

author
#0 lore raw

PATCH v2 1/1 crypto: xor - fix template benchmarking

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 machines 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 shows up 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 <[email protected]>
Tested-by John David Anglin <[email protected]>

v2:
- clean up coding style (noticed & suggested by Herbert Xu)
- rephrased & fixed typo in commit message

crypto/xor.c +14 -17
index 8e72e5d5db0d..56aa3169e871 100644
@@ -83,33 +83,30 @@ 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;
-	}
+	reps = 0;
+	t0 = ktime_get();
+	/* delay start until time has advanced */
+	while ((start = ktime_get()) == t0)
+		cpu_relax();
+	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);
maintainer
#1 lore raw
Quoted 27 lines · Helge Deller

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 machines 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 shows up 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.

Quoted 7 lines
Fixes c055e3eae0f1 ("crypto: xor - use ktime for template benchmarking")
Signed-off-by Helge Deller <[email protected]>
Tested-by John David Anglin <[email protected]>

v2:
- clean up coding style (noticed & suggested by Herbert Xu)
- rephrased & fixed typo in commit message

Patch applied. Thanks.

Signature · 3 lines
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt