LTTng relay-alloc cache tests
With these tests, I come to the surprising conclusion that it is faster to keep the buffer pages as write-back than to set them to write-combined or uncached. I would have thought that keeping the cachelines free for kernel or userspace cache lines would have improved performance, but for some reason it doesn't.
Doing these tests under a webserver-like workload would however be interesting. Feel free to play with this. :-)
Mathieu
LTTng trace buffer memory cache behavior comparison Tracing in flight recorder mode with default size buffers LTTng 0.31, kernel 2.6.27-rc8
Dual quad-core x86_64 with 16GB ram
** tbench (MB/s) **
tracing disabled : 1862.24
tracing flight recorder default buffer size :
cache WB : 942.72 cache WC (write-combined) : 755.94 uncached : 780.03
** kernel compilation **
make clean
sync
sync
echo 3 > /proc/sys/vm/drop_caches
time make -j10
[timing 1, cache cold]
make clean
make -j10
[timing 2, cache hot]
Tracing disabled :
[timing 1, cache cold]
real 1m24.729s user 7m22.952s sys 0m48.943s
[timing 2, cache hot]
real 1m10.639s user 7m25.864s sys 0m49.235s
cache WB :
[timing 1, cache cold]
real 1m33.408s user 7m27.852s sys 1m47.623s
[timing 2, cache hot]
real 1m20.429s user 7m30.668s sys 1m46.895s
cache WC (write-combined) :
[timing 1, cache cold]
real 1m38.971s user 7m27.800s sys 2m18.241s
[timing 2, cache hot]
real 1m25.200s user 7m30.040s sys 2m18.201s
uncached :
[timing 1, cache cold]
real 1m37.237s user 7m28.352s sys 2m13.312s
[timing 2, cache hot]
real 1m24.303s user 7m29.420s sys 2m14.756s
--- ltt/ltt-relay-alloc.c | 40 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 40 insertions(+)
ltt/ltt-relay-alloc.c +40 -0
@@ -26,6 +26,38 @@static DEFINE_MUTEX (relay_channels_mutex);static LIST_HEAD (relay_channels); +#ifdef CONFIG_X86 +#include <asm/cacheflush.h> +#include <asm/pat.h> + +/* +* TODO : create an abstraction in arch/x86 +* Use write-combined pages if PAT is supported. Else, use uncached pages to +* make sure we don't evince really useful cachelines. +*/ +static int arch_buffer_page_attr_set (struct page *page,int numpages) +{ +return set_memory_wc ((unsigned long )page_address (page), numpages); +//return set_memory_uc((unsigned long)page_address(page), numpages); +} + +static int arch_buffer_page_attr_clear (struct page *page,int numpages) +{ +return set_memory_wb ((unsigned long )page_address (page), numpages); +} +#else +static int arch_buffer_page_attr_set (struct page *page) +{ +return 0 ; +} + +static int arch_buffer_page_attr_clear (struct page *page) +{ +return 0 ; +} +#endif + +/** * relay_alloc_buf - allocate a channel buffer * @buf: the buffer struct
@@ -35,6 +67,7 @@ static int relay_alloc_buf(struct rchan_ {unsigned int i, n_pages;struct buf_page *buf_page, *n; +int ret; *size =PAGE_ALIGN (*size); n_pages = *size >>PAGE_SHIFT ;
@@ -56,6 +89,8 @@ static int relay_alloc_buf(struct rchan_ buf_page->offset = (size_t )i <<PAGE_SHIFT ; buf_page->buf = buf;set_page_private (buf_page->page, (unsigned long )buf_page); + ret =arch_buffer_page_attr_set (buf_page->page,1 ); +WARN_ON (ret);if (i ==0 ) { buf->wpage = buf_page; buf->hpage[0 ] = buf_page;
@@ -69,6 +104,8 @@ static int relay_alloc_buf(struct rchan_ depopulate:list_for_each_entry_safe (buf_page, n, &buf->pages, list) {list_del_init (&buf_page->list); + ret =arch_buffer_page_attr_clear (buf_page->page,1 ); +WARN_ON (ret);__free_page (buf_page->page);kfree (buf_page); }
@@ -123,9 +160,12 @@ static void relay_destroy_buf(struct rch {struct rchan *chan = buf->chan;struct buf_page *buf_page, *n; +int ret;list_for_each_entry_safe (buf_page, n, &buf->pages, list) {list_del_init (&buf_page->list); + ret =arch_buffer_page_attr_clear (buf_page->page,1 ); +WARN_ON (ret);__free_page (buf_page->page);kfree (buf_page); }
Signature · 2 lines
Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68