Skip to contentSkip to conversation

selftests/vm: 8x compaction_test speedup

PATCH Applied John Hubbard 3 participants 4 messages Oct 2, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover selftests/vm: 8x compaction_test speedup

Hi,

Maybe this should really be an RFC, given that I don't fully understand why the compaction_test.c program was mmap'ing 1 MB at a time. So apologies in advance if I've mucked up something important, but if so, maybe we can still find a way to get this fixed up to something better.

Anyway: there are 20+ tests in tools/testing/selftests/vm/. The entire running time for these (via run_vmtest.sh) is about 56 seconds, of which over half is due to just one test: compaction_test, which takes 27 sec! (A runner-up is HMM, at 11 sec, so it's up for a look next.) The other tests mostly take a few ms, and a few take 1.0 sec.

This drops the compaction_test run time from 27, to 3.3 sec. Enjoy. :)

thanks,
John Hubbard
NVIDIA

John Hubbard (1):
selftests/vm: 8x compaction_test speedup

 tools/testing/selftests/vm/compaction_test.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
Signature · 1 line
2.28.0
#1 lore raw

PATCH 1/1 selftests/vm: 8x compaction_test speedup

This patch reduces the running time for compaction_test from about 27 sec, to 3.3 sec, which is about an 8x speedup.

These numbers are for an Intel x86_64 system with 32 GB of DRAM.

The compaction_test.c program was spending most of its time doing mmap(), 1 MB at a time, on about 25 GB of memory.

Instead, do the mmaps 100 MB at a time. (Going past 100 MB doesn't make things go much faster, because other parts of the program are using the remaining time.)

---
 tools/testing/selftests/vm/compaction_test.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
tools/testing/selftests/vm/compaction_test.c +6 -5
index bcec71250873..9b420140ba2b 100644
@@ -18,7 +18,8 @@
 
 #include "../kselftest.h"
 
-#define MAP_SIZE 1048576
+#define MAP_SIZE_MB	100
+#define MAP_SIZE	(MAP_SIZE_MB * 1024 * 1024)
 
 struct map_list {
 	void *map;
@@ -165,7 +166,7 @@ int main(int argc, char **argv)
 	void *map = NULL;
 	unsigned long mem_free = 0;
 	unsigned long hugepage_size = 0;
-	unsigned long mem_fragmentable = 0;
+	long mem_fragmentable_MB = 0;
 
 	if (prereq() != 0) {
 		printf("Either the sysctl compact_unevictable_allowed is not\n"
@@ -190,9 +191,9 @@ int main(int argc, char **argv)
 		return -1;
 	}
 
-	mem_fragmentable = mem_free * 0.8 / 1024;
+	mem_fragmentable_MB = mem_free * 0.8 / 1024;
 
-	while (mem_fragmentable > 0) {
+	while (mem_fragmentable_MB > 0) {
 		map = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
 			   MAP_ANONYMOUS | MAP_PRIVATE | MAP_LOCKED, -1, 0);
 		if (map == MAP_FAILED)
@@ -213,7 +214,7 @@ int main(int argc, char **argv)
 		for (i = 0; i < MAP_SIZE; i += page_size)
 			*(unsigned long *)(map + i) = (unsigned long)map + i;
 
-		mem_fragmentable--;
+		mem_fragmentable_MB -= MAP_SIZE_MB;
 	}
 
 	for (entry = list; entry != NULL; entry = entry->next) {
Signature · 1 line
2.28.0
maintainer
#2 lore raw
Quoted 11 lines · On Fri, 2 Oct 2020 01:06:21 -0700 John Hubbard <[email protected]> wrote:

This patch reduces the running time for compaction_test from about 27 sec, to 3.3 sec, which is about an 8x speedup.

These numbers are for an Intel x86_64 system with 32 GB of DRAM.

The compaction_test.c program was spending most of its time doing mmap(), 1 MB at a time, on about 25 GB of memory.

Instead, do the mmaps 100 MB at a time. (Going past 100 MB doesn't make things go much faster, because other parts of the program are using the remaining time.)

Seems nice. It's been 5 years, but hopefully Sri is still at Akamai?

Quoted 45 lines
tools/testing/selftests/vm/compaction_test.c +6 -5
@@ -18,7 +18,8 @@
 
 #include "../kselftest.h"
 
-#define MAP_SIZE 1048576
+#define MAP_SIZE_MB	100
+#define MAP_SIZE	(MAP_SIZE_MB * 1024 * 1024)
 
 struct map_list {
 	void *map;
@@ -165,7 +166,7 @@ int main(int argc, char **argv)
 	void *map = NULL;
 	unsigned long mem_free = 0;
 	unsigned long hugepage_size = 0;
-	unsigned long mem_fragmentable = 0;
+	long mem_fragmentable_MB = 0;
 
 	if (prereq() != 0) {
 		printf("Either the sysctl compact_unevictable_allowed is not\n"
@@ -190,9 +191,9 @@ int main(int argc, char **argv)
 		return -1;
 	}
 
-	mem_fragmentable = mem_free * 0.8 / 1024;
+	mem_fragmentable_MB = mem_free * 0.8 / 1024;
 
-	while (mem_fragmentable > 0) {
+	while (mem_fragmentable_MB > 0) {
 		map = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
 			   MAP_ANONYMOUS | MAP_PRIVATE | MAP_LOCKED, -1, 0);
 		if (map == MAP_FAILED)
@@ -213,7 +214,7 @@ int main(int argc, char **argv)
 		for (i = 0; i < MAP_SIZE; i += page_size)
 			*(unsigned long *)(map + i) = (unsigned long)map + i;
 
-		mem_fragmentable--;
+		mem_fragmentable_MB -= MAP_SIZE_MB;
 	}
 
 	for (entry = list; entry != NULL; entry = entry->next) {
Signature · 1 line
2.28.0
#3 lore raw

On 10/2/20, 6:05 PM, "Andrew Morton" <[email protected]> wrote:

On Fri, 2 Oct 2020 01:06:21 -0700 John Hubbard <[email protected]> wrote:

> This patch reduces the running time for compaction_test from about 27 > sec, to 3.3 sec, which is about an 8x speedup.
>
> These numbers are for an Intel x86_64 system with 32 GB of DRAM. >
> The compaction_test.c program was spending most of its time doing > mmap(), 1 MB at a time, on about 25 GB of memory.
>
> Instead, do the mmaps 100 MB at a time. (Going past 100 MB doesn't make > things go much faster, because other parts of the program are using the > remaining time.)

Seems nice. It's been 5 years, but hopefully Sri is still at Akamai?

Looks good to me.

Thanks
-Sri