Date:   Mon,  7 Aug 2017 13:40:33 +0800
From:   "Huang, Ying" <ying.huang@...el.com>
To:     Andrew Morton <akpm@...ux-foundation.org>
Cc:     linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        "Huang, Ying" <ying.huang@...el.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Minchan Kim <minchan@...nel.org>,
        Rik van Riel <riel@...hat.com>, Shaohua Li <shli@...nel.org>,
        Hugh Dickins <hughd@...gle.com>,
        Fengguang Wu <fengguang.wu@...el.com>,
        Tim Chen <tim.c.chen@...el.com>,
        Dave Hansen <dave.hansen@...el.com>
Subject: [PATCH -mm -v4 0/5] mm, swap: VMA based swap readahead

The swap readahead is an important mechanism to reduce the swap in
latency.  Although pure sequential memory access pattern isn't very
popular for anonymous memory, the space locality is still considered
valid.

In the original swap readahead implementation, the consecutive blocks
in swap device are readahead based on the global space locality
estimation.  But the consecutive blocks in swap device just reflect
the order of page reclaiming, don't necessarily reflect the access
pattern in virtual memory space.  And the different tasks in the
system may have different access patterns, which makes the global
space locality estimation incorrect.

In this patchset, when page fault occurs, the virtual pages near the
fault address will be readahead instead of the swap slots near the
fault swap slot in swap device.  This avoid to readahead the unrelated
swap slots.  At the same time, the swap readahead is changed to work
on per-VMA from globally.  So that the different access patterns of
the different VMAs could be distinguished, and the different readahead
policy could be applied accordingly.  The original core readahead
detection and scaling algorithm is reused, because it is an effect
algorithm to detect the space locality.

In addition to the swap readahead changes, some new sysfs interface is
added to show the efficiency of the readahead algorithm and some other
swap statistics.

This new implementation will incur more small random read, on SSD, the
improved correctness of estimation and readahead target should beat
the potential increased overhead, this is also illustrated in the test
results below.  But on HDD, the overhead may beat the benefit, so the
original implementation will be used by default.

The test and result is as follow,

Common test condition
=====================

Test Machine: Xeon E5 v3 (2 sockets, 72 threads, 32G RAM)
Swap device: NVMe disk

Micro-benchmark with combined access pattern
============================================

vm-scalability, sequential swap test case, 4 processes to eat 50G
virtual memory space, repeat the sequential memory writing until 300
seconds.  The first round writing will trigger swap out, the following
rounds will trigger sequential swap in and out.

At the same time, run vm-scalability random swap test case in
background, 8 processes to eat 30G virtual memory space, repeat the
random memory write until 300 seconds.  This will trigger random
swap-in in the background.

This is a combined workload with sequential and random memory
accessing at the same time.  The result (for sequential workload) is
as follow,

			Base		Optimized
			----		---------
throughput		345413 KB/s	414029 KB/s (+19.9%)
latency.average		97.14 us	61.06 us (-37.1%)
latency.50th		2 us		1 us
latency.60th		2 us		1 us
latency.70th		98 us		2 us
latency.80th		160 us		2 us
latency.90th		260 us		217 us
latency.95th		346 us		369 us
latency.99th		1.34 ms		1.09 ms
ra_hit%			52.69%		99.98%

The original swap readahead algorithm is confused by the background
random access workload, so readahead hit rate is lower.  The VMA-base
readahead algorithm works much better.

Linpack
=======

The test memory size is bigger than RAM to trigger swapping.

			Base		Optimized
			----		---------
elapsed_time		393.49 s	329.88 s (-16.2%)
ra_hit%			86.21%		98.82%

The score of base and optimized kernel hasn't visible changes.  But
the elapsed time reduced and readahead hit rate improved, so the
optimized kernel runs better for startup and tear down stages.  And
the absolute value of readahead hit rate is high, shows that the space
locality is still valid in some practical workloads.

Changelogs:

v4:

- Rebased on latest -mm tree.

- Remove swap cache statistics interface, because we found that the
  interface for readahead statistics should be sufficient.

- Use /proc/vmstat for swap readahead statistics, because that is the
  interface used by other similar statistics.

- Add ABI document for newly added sysfs interface.

v3:

- Rebased on latest -mm tree

- Use percpu_counter for swap readahead statistics per Dave Hansen's comment.

Best Regards,
Huang, Ying

Powered by blists - more mailing lists