Content-Type: multipart/mixed; boundary="===============2297643107643905305==" MIME-Version: 1.0 From: Peter Hurley To: lkp@lists.01.org Subject: Re: increased vmap_area_lock contentions on "n_tty: Move buffers into n_tty_data" Date: Tue, 17 Sep 2013 20:49:51 -0400 Message-ID: <5238F8AF.4050309@hurleysoftware.com> In-Reply-To: <20130913010936.GA1291@localhost> List-Id: --===============2297643107643905305== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 09/12/2013 09:09 PM, Fengguang Wu wrote: > On Fri, Sep 13, 2013 at 08:51:33AM +0800, Fengguang Wu wrote: >> Hi Peter, >> >> FYI, we noticed much increased vmap_area_lock contentions since this >> commit: >> >> commit 20bafb3d23d108bc0a896eb8b7c1501f4f649b77 >> Author: Peter Hurley >> Date: Sat Jun 15 10:21:19 2013 -0400 >> >> n_tty: Move buffers into n_tty_data >> >> Reduce pointer reloading and improve locality-of-reference; >> allocate read_buf and echo_buf within struct n_tty_data. > > Here are some comparison between this commit [o] with its parent commit [= *]. [...] > 8cb06c983822103da1cf 20bafb3d23d108bc0a89 > ------------------------ ------------------------ > 976.67 +108.3% 2034.67 lkp-a04/micro/netper= f/120s-200%-TCP_STREAM > 8971.36 +11.4% 9997.05 nhm-white/micro/aim7= /exec_test > 9948.03 +20.9% 12031.72 TOTAL slabinfo.kmall= oc-128.active_objs > > 8cb06c983822103da1cf 20bafb3d23d108bc0a89 > ------------------------ ------------------------ > 976.67 +108.3% 2034.67 lkp-a04/micro/netper= f/120s-200%-TCP_STREAM > 9127.64 +11.4% 10164.15 nhm-white/micro/aim7= /exec_test > 10104.31 +20.7% 12198.82 TOTAL slabinfo.kmall= oc-128.num_objs The dramatic increase in 128-byte kmalloc blocks is from vmalloc overhead with associated with each allocation. On a x86_64, struct vmap_area is 104 = bytes, rounded to 128, allocated with every vmalloc allocation. This is approx 1% = overhead (which seems high to me). The reason this is still visible after the test completes is the vmap area = is lazily reclaimed (see mm/vmalloc.c:__purge_vmap_area_lazy()). 1% memory overhead coupled with the unwanted vmap_area_lock contention (eve= n though it is test-induced) -- I might revert this anyway. Regards, Peter Hurley --===============2297643107643905305==--