object allocation benchmark

I was wondering: Which benchmark would be considered the canonical one to demonstrate the speed of the slub/slab after changes? In particular, I have the kmem-memcg in mind

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I have some in kernel benchmarking tools for page allocator and slab allocators. But they are not really clean patches.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

This is the latest version of your tools that I have based on 3.3. Load the modules with insmod and it will produce an error to automatically unloaded (by design) and check dmesg for the results.

---
 Makefile               |    2 +-
 include/Kbuild         |    1 +
 lib/Kconfig.debug      |    1 +
 tests/Kconfig          |   32 +++++
 tests/Makefile         |    3 +
 tests/pagealloc_test.c |  334 +++++++++++++++++++++++++++++++++++++++++++
 tests/slab_test.c      |  372 ++++++++++++++++++++++++++++++++++++++++++++++++
 tests/vmstat_test.c    |   96 +++++++++++++
 8 files changed, 840 insertions(+), 1 deletion(-)
 create mode 100644 tests/Kconfig
 create mode 100644 tests/Makefile
 create mode 100644 tests/pagealloc_test.c
 create mode 100644 tests/slab_test.c
 create mode 100644 tests/vmstat_test.c

@@ -708,7 +708,7 @@ export mod_strip_cmd
 
 
 ifeq ($(KBUILD_EXTMOD),)
-core-y		+= kernel/ mm/ fs/ ipc/ security/ crypto/ block/
+core-y		+= kernel/ mm/ fs/ ipc/ security/ crypto/ block/ tests/
 
 vmlinux-dirs	:= $(patsubst %/,%,$(filter %/, $(init-y) $(init-m) \
 		     $(core-y) $(core-m) $(drivers-y) $(drivers-m) \

@@ -10,3 +10,4 @@ header-y += video/
 header-y += drm/
 header-y += xen/
 header-y += scsi/
+header-y += tests/

@@ -1123,6 +1123,7 @@ config SYSCTL_SYSCALL_CHECK
 
 source mm/Kconfig.debug
 source kernel/trace/Kconfig
+source tests/Kconfig
 
 config PROVIDE_OHCI1394_DMA_INIT
 	bool "Remote debugging over FireWire early on boot"

new file mode 100644
@@ -0,0 +1,32 @@
+menuconfig BENCHMARKS
+	bool "In kernel benchmarks"
+	def_bool n
+	help
+	  Includes in kernel benchmark modules in the build. These modules can
+	  be loaded later to trigger benchmarking kernel subsystems.
+	  Output will be generated in the system log.
+
+if BENCHMARKS
+
+config BENCHMARK_SLAB
+	tristate "Slab allocator Benchmark"
+	depends on m
+	default m
+	help
+	  A benchmark that measures slab allocator performance.
+
+config BENCHMARK_VMSTAT
+	tristate "VM statistics Benchmark"
+	depends on m
+	default m
+	help
+	  A benchmark measuring the performance of vm statistics.
+
+config BENCHMARK_PAGEALLOC
+	tristate "Page Allocator Benchmark"
+	depends on m
+	default m
+	help
+	  A benchmark measuring the performance of the page allocator.
+
+endif # BENCHMARKS

new file mode 100644
@@ -0,0 +1,3 @@
+obj-$(CONFIG_BENCHMARK_SLAB) += slab_test.o
+obj-$(CONFIG_BENCHMARK_VMSTAT) += vmstat_test.o
+obj-$(CONFIG_BENCHMARK_PAGEALLOC) += pagealloc_test.o

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Anyone interested in pushing the benchmark to mainline?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

It's a pretty atypical type of a benchmark that must be compiled and loaded as a benchmark for the most accurate results, I'm not sure if we want to carry it in the kernel or not. I got these from when Christoph initially sent them for your tree. If there's a renewed interest, let me know and I'll send you the ported changes (pagealloc_test.c, slab_test.c, and vmstat_test.c).

I'd given it a try.

So in general, Suleiman patches perform fine against bare slab, the differences being in the order of ~ 1%. There are some spikes a little bit above that, that would deserve more analysis.

However, reason I decided to report early, is this test:
"1 alloc N free test". It is quite erratic. memcg+kmem sometimes performs 15 % worse, sometimes 30 % better... Always right after a cold boot.

I was wondering if you usually see such behavior for this test, and has some tips on the setup in case I'm doing anything wrong ?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Well some of these tests are sensitive to memory fragmentation in the page allocator and slab allocator. One approach is to do these tests immediately after boot with minimal user space bringup.

The other is to run a stress tests for an predefined period to get memory into a sufficiently fragmented state and then run the test.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>