mm/vmalloc: Introduce DEBUG_VMALLOCINFO to reduce spinlock contention
I applied this patch to my kernel source (3.13-based) and everything seems to work smoother
anyone else wanna try ?
we need a news entry about this on phoronix !
every lock removed - towards a lockless state - improves responsiveness and cuts down latency
Performance analysis of software compilation by Gentoo portage on an
Intel E5-2620 with 64GB of RAM revealed that a sizeable amount of time,
anywhere from 5% to 15%, was spent in get_vmalloc_info(), with at least
40% of that time spent in the _raw_spin_lock() invoked by it.
The spinlock call is done on vmap_area_lock to protect vmap_area_list,
but changes to vmap_area_list are made under RCU. The only consumer that
requires a spinlock on an RCU-ified list is /proc/vmallocinfo. That is
only intended for use by kernel developers doing debugging, but even few
kernel developers appear to use it. Introducing DEBUG_VMALLOCINFO allows
us to fully RCU-ify the list, which eliminates this list as a source of
contention.
This patch brings a substantial reduction in time spent in spinlocks on
my system. Flame graphs from my early analysis are available on my
developer space. They were created by profiling the system under
concurrent package builds done by emerge at a sample rate of 99Hz for 10
seconds and using Brendan Gregg's scripts to process the data:
In this example, 6.64% of system time is spent in get_vmalloc_info()
with 2.59% spent in the spinlock. The patched version sees only 0.50% of
time spent in get_vmalloc_info() with neligible time spent in spin
locks. The low utilization of get_vmalloc_info() in this is partly
attributable to measurement error, but the reduction in time spent
spinning is clear.
Intel E5-2620 with 64GB of RAM revealed that a sizeable amount of time,
anywhere from 5% to 15%, was spent in get_vmalloc_info(), with at least
40% of that time spent in the _raw_spin_lock() invoked by it.
The spinlock call is done on vmap_area_lock to protect vmap_area_list,
but changes to vmap_area_list are made under RCU. The only consumer that
requires a spinlock on an RCU-ified list is /proc/vmallocinfo. That is
only intended for use by kernel developers doing debugging, but even few
kernel developers appear to use it. Introducing DEBUG_VMALLOCINFO allows
us to fully RCU-ify the list, which eliminates this list as a source of
contention.
This patch brings a substantial reduction in time spent in spinlocks on
my system. Flame graphs from my early analysis are available on my
developer space. They were created by profiling the system under
concurrent package builds done by emerge at a sample rate of 99Hz for 10
seconds and using Brendan Gregg's scripts to process the data:
In this example, 6.64% of system time is spent in get_vmalloc_info()
with 2.59% spent in the spinlock. The patched version sees only 0.50% of
time spent in get_vmalloc_info() with neligible time spent in spin
locks. The low utilization of get_vmalloc_info() in this is partly
attributable to measurement error, but the reduction in time spent
spinning is clear.
anyone else wanna try ?
we need a news entry about this on phoronix !
every lock removed - towards a lockless state - improves responsiveness and cuts down latency
Comment