Hi,
Currently, m_tag_get() calls malloc()(kern_malloc()), and then
kern_malloc() calls pool_cache_get() as long as m_tag size is not too
large.
According to my measurements of using DTrace, there is a certain
overhead between kern_malloc() and pool_cache_get(). Here is each
function's average turnaround times in my measurement.
m_tag_get | 124205 [ns]
kern_malloc | 108007 [ns]
kmem_intr_alloc | 89636 [ns]
pool_cache_get_paddr | 65942 [ns]
i.e. the overhead between kern_malloc() and pool_cache_get() is 42065[ns],
this overhead is considerable compared to pool_cache_get_paddr()'s
turnaround time.
Of course, m_tag_get() is used by packet processing path, this overhead
would have a certain influence on packet throughput and latency.
So, I think m_tag should use a dedicated pool cache like mbuf to call
pool_cache_get() directly.
Here is the patch to introduce mbuf pool cache.
http://www.netbsd.org/~knakahara/mtag-pool-cache/mtag-pool-cache.patch
Could you comment this patch? Any comments are welcome.
Thanks,
--
//////////////////////////////////////////////////////////////////////
Internet Initiative Japan Inc.
Device Engineering Section,
Core Product Development Department,
Product Division,
Technology Unit
Kengo NAKAHARA <k-nakah...@iij.ad.jp>
