mm/ksm: add ksm advisor

Patch series "mm/ksm: Add ksm advisor", v5.

What is the KSM advisor?
=========================
The ksm advisor automatically manages the pages_to_scan setting to achieve
a target scan time.  The target scan time defines how many seconds it
should take to scan all the candidate KSM pages.  In other words the
pages_to_scan rate is changed by the advisor to achieve the target scan
time.

Why do we need a KSM advisor?
==============================
The number of candidate pages for KSM is dynamic.  It can often be
observed that during the startup of an application more candidate pages
need to be processed.  Without an advisor the pages_to_scan parameter
needs to be sized for the maximum number of candidate pages.  With the
scan time advisor the pages_to_scan parameter based can be changed based
on demand.

Algorithm
==========
The algorithm calculates the change value based on the target scan time
and the previous scan time.  To avoid pertubations an exponentially
weighted moving average is applied.

The algorithm has a max and min
value to:
- guarantee responsiveness to changes
- to limit CPU resource consumption

Parameters to influence the KSM scan advisor
=============================================
The respective parameters are:
- ksm_advisor_mode
  0: None (default), 1: scan time advisor
- ksm_advisor_target_scan_time
  how many seconds a scan should of all candidate pages take
- ksm_advisor_max_cpu
  upper limit for the cpu usage in percent of the ksmd background thread

The initial value and the max value for the pages_to_scan parameter can
be limited with:
- ksm_advisor_min_pages_to_scan
  minimum value for pages_to_scan per batch
- ksm_advisor_max_pages_to_scan
  maximum value for pages_to_scan per batch

The default settings for the above two parameters should be suitable for
most workloads.

The parameters are exposed as knobs in /sys/kernel/mm/ksm. By default the
scan time advisor is disabled.

Currently there are two advisors:
- none and
- scan-time.

Resource savings
=================
Tests with various workloads have shown considerable CPU savings. Most
of the workloads I have investigated have more candidate pages during
startup. Once the workload is stable in terms of memory, the number of
candidate pages is reduced. Without the advisor, the pages_to_scan needs
to be sized for the maximum number of candidate pages. So having this
advisor definitely helps in reducing CPU consumption.

For the instagram workload, the advisor achieves a 25% CPU reduction.
Once the memory is stable, the pages_to_scan parameter gets reduced to
about 40% of its max value.

The new advisor works especially well if the smart scan feature is also
enabled.

How is defining a target scan time better?
===========================================
For an administrator it is more logical to set a target scan time.. The
administrator can determine how many pages are scanned on each scan.
Therefore setting a target scan time makes more sense.

In addition the administrator might have a good idea about the memory
sizing of its respective workloads.

Setting cpu limits is easier than setting The pages_to_scan parameter. The
pages_to_scan parameter is per batch. For the administrator it is difficult
to set the pages_to_scan parameter.

Tracing
=======
A new tracing event has been added for the scan time advisor. The new
trace event is called ksm_advisor. It reports the scan time, the new
pages_to_scan setting and the cpu usage of the ksmd background thread.

Other approaches
=================

Approach 1: Adapt pages_to_scan after processing each batch. If KSM
  merges pages, increase the scan rate, if less KSM pages, reduce the
  the pages_to_scan rate. This doesn't work too well. While it increases
  the pages_to_scan for a short period, but generally it ends up with a
  too low pages_to_scan rate.

Approach 2: Adapt pages_to_scan after each scan. The problem with that
  approach is that the calculated scan rate tends to be high. The more
  aggressive KSM scans, the more pages it can de-duplicate.

There have been earlier attempts at an advisor:
  propose auto-run mode of ksm and its tests
  (https://marc.info/?l=linux-mm&m=166029880214485&w=2)

This patch (of 5):

This adds the ksm advisor.  The ksm advisor automatically manages the
pages_to_scan setting to achieve a target scan time.  The target scan time
defines how many seconds it should take to scan all the candidate KSM
pages.  In other words the pages_to_scan rate is changed by the advisor to
achieve the target scan time.  The algorithm has a max and min value to:

- guarantee responsiveness to changes
- limit CPU resource consumption

The respective parameters are:
- ksm_advisor_target_scan_time (how many seconds a scan should take)
- ksm_advisor_max_cpu (maximum value for cpu percent usage)

- ksm_advisor_min_pages (minimum value for pages_to_scan per batch)
- ksm_advisor_max_pages (maximum value for pages_to_scan per batch)

The algorithm calculates the change value based on the target scan time
and the previous scan time. To avoid pertubations an exponentially
weighted moving average is applied.

The advisor is managed by two main parameters: target scan time,
cpu max time for the ksmd background thread. These parameters determine
how aggresive ksmd scans.

In addition there are min and max values for the pages_to_scan parameter
to make sure that its initial and max values are not set too low or too
high.  This ensures that it is able to react to changes quickly enough.

The default values are:
- target scan time: 200 secs
- max cpu: 70%
- min pages: 500
- max pages: 30000

By default the advisor is disabled. Currently there are two advisors:
none and scan-time.

Tests with various workloads have shown considerable CPU savings.  Most of
the workloads I have investigated have more candidate pages during
startup, once the workload is stable in terms of memory, the number of
candidate pages is reduced.  Without the advisor, the pages_to_scan needs
to be sized for the maximum number of candidate pages.  So having this
advisor definitely helps in reducing CPU consumption.

For the instagram workload, the advisor achieves a 25% CPU reduction.
Once the memory is stable, the pages_to_scan parameter gets reduced to
about 40% of its max value.

Link: https://lkml.kernel.org/r/20231218231054.1625219-1-shr@devkernel.io
Link: https://lkml.kernel.org/r/20231218231054.1625219-2-shr@devkernel.io
Signed-off-by: Stefan Roesch <shr@devkernel.io>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Stefan Roesch <shr@devkernel.io>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

diff --git a/mm/ksm.c b/mm/ksm.c
index e2ce850c27391d6dd48e159ae03940e5924ad0db..599f7d209c68d18ab4af5b5f1bf32df4ce91bd57 100644 (file)
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -21,6 +21,7 @@
 #include <linux/sched.h>
 #include <linux/sched/mm.h>
 #include <linux/sched/coredump.h>
+#include <linux/sched/cputime.h>
 #include <linux/rwsem.h>
 #include <linux/pagemap.h>
 #include <linux/rmap.h>
@@ -248,6 +249,9 @@ static struct kmem_cache *rmap_item_cache;
 static struct kmem_cache *stable_node_cache;
 static struct kmem_cache *mm_slot_cache;
 
+/* Default number of pages to scan per batch */
+#define DEFAULT_PAGES_TO_SCAN 100
+
 /* The number of pages scanned */
 static unsigned long ksm_pages_scanned;
 
@@ -276,7 +280,7 @@ static unsigned int ksm_stable_node_chains_prune_millisecs = 2000;
 static int ksm_max_page_sharing = 256;
 
 /* Number of pages ksmd should scan in one batch */
-static unsigned int ksm_thread_pages_to_scan = 100;
+static unsigned int ksm_thread_pages_to_scan = DEFAULT_PAGES_TO_SCAN;
 
 /* Milliseconds ksmd should sleep between batches */
 static unsigned int ksm_thread_sleep_millisecs = 20;
@@ -297,6 +301,152 @@ unsigned long ksm_zero_pages;
 /* The number of pages that have been skipped due to "smart scanning" */
 static unsigned long ksm_pages_skipped;
 
+/* Don't scan more than max pages per batch. */
+static unsigned long ksm_advisor_max_pages_to_scan = 30000;
+
+/* Min CPU for scanning pages per scan */
+#define KSM_ADVISOR_MIN_CPU 10
+
+/* Max CPU for scanning pages per scan */
+static unsigned int ksm_advisor_max_cpu =  70;
+
+/* Target scan time in seconds to analyze all KSM candidate pages. */
+static unsigned long ksm_advisor_target_scan_time = 200;
+
+/* Exponentially weighted moving average. */
+#define EWMA_WEIGHT 30
+
+/**
+ * struct advisor_ctx - metadata for KSM advisor
+ * @start_scan: start time of the current scan
+ * @scan_time: scan time of previous scan
+ * @change: change in percent to pages_to_scan parameter
+ * @cpu_time: cpu time consumed by the ksmd thread in the previous scan
+ */
+struct advisor_ctx {
+       ktime_t start_scan;
+       unsigned long scan_time;
+       unsigned long change;
+       unsigned long long cpu_time;
+};
+static struct advisor_ctx advisor_ctx;
+
+/* Define different advisor's */
+enum ksm_advisor_type {
+       KSM_ADVISOR_NONE,
+       KSM_ADVISOR_SCAN_TIME,
+};
+static enum ksm_advisor_type ksm_advisor;
+
+static inline void advisor_start_scan(void)
+{
+       if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
+               advisor_ctx.start_scan = ktime_get();
+}
+
+/*
+ * Use previous scan time if available, otherwise use current scan time as an
+ * approximation for the previous scan time.
+ */
+static inline unsigned long prev_scan_time(struct advisor_ctx *ctx,
+                                          unsigned long scan_time)
+{
+       return ctx->scan_time ? ctx->scan_time : scan_time;
+}
+
+/* Calculate exponential weighted moving average */
+static unsigned long ewma(unsigned long prev, unsigned long curr)
+{
+       return ((100 - EWMA_WEIGHT) * prev + EWMA_WEIGHT * curr) / 100;
+}
+
+/*
+ * The scan time advisor is based on the current scan rate and the target
+ * scan rate.
+ *
+ *      new_pages_to_scan = pages_to_scan * (scan_time / target_scan_time)
+ *
+ * To avoid perturbations it calculates a change factor of previous changes.
+ * A new change factor is calculated for each iteration and it uses an
+ * exponentially weighted moving average. The new pages_to_scan value is
+ * multiplied with that change factor:
+ *
+ *      new_pages_to_scan *= change facor
+ *
+ * The new_pages_to_scan value is limited by the cpu min and max values. It
+ * calculates the cpu percent for the last scan and calculates the new
+ * estimated cpu percent cost for the next scan. That value is capped by the
+ * cpu min and max setting.
+ *
+ * In addition the new pages_to_scan value is capped by the max and min
+ * limits.
+ */
+static void scan_time_advisor(void)
+{
+       unsigned int cpu_percent;
+       unsigned long cpu_time;
+       unsigned long cpu_time_diff;
+       unsigned long cpu_time_diff_ms;
+       unsigned long pages;
+       unsigned long per_page_cost;
+       unsigned long factor;
+       unsigned long change;
+       unsigned long last_scan_time;
+       unsigned long scan_time;
+
+       /* Convert scan time to seconds */
+       scan_time = div_s64(ktime_ms_delta(ktime_get(), advisor_ctx.start_scan),
+                           MSEC_PER_SEC);
+       scan_time = scan_time ? scan_time : 1;
+
+       /* Calculate CPU consumption of ksmd background thread */
+       cpu_time = task_sched_runtime(current);
+       cpu_time_diff = cpu_time - advisor_ctx.cpu_time;
+       cpu_time_diff_ms = cpu_time_diff / 1000 / 1000;
+
+       cpu_percent = (cpu_time_diff_ms * 100) / (scan_time * 1000);
+       cpu_percent = cpu_percent ? cpu_percent : 1;
+       last_scan_time = prev_scan_time(&advisor_ctx, scan_time);
+
+       /* Calculate scan time as percentage of target scan time */
+       factor = ksm_advisor_target_scan_time * 100 / scan_time;
+       factor = factor ? factor : 1;
+
+       /*
+        * Calculate scan time as percentage of last scan time and use
+        * exponentially weighted average to smooth it
+        */
+       change = scan_time * 100 / last_scan_time;
+       change = change ? change : 1;
+       change = ewma(advisor_ctx.change, change);
+
+       /* Calculate new scan rate based on target scan rate. */
+       pages = ksm_thread_pages_to_scan * 100 / factor;
+       /* Update pages_to_scan by weighted change percentage. */
+       pages = pages * change / 100;
+
+       /* Cap new pages_to_scan value */
+       per_page_cost = ksm_thread_pages_to_scan / cpu_percent;
+       per_page_cost = per_page_cost ? per_page_cost : 1;
+
+       pages = min(pages, per_page_cost * ksm_advisor_max_cpu);
+       pages = max(pages, per_page_cost * KSM_ADVISOR_MIN_CPU);
+       pages = min(pages, ksm_advisor_max_pages_to_scan);
+
+       /* Update advisor context */
+       advisor_ctx.change = change;
+       advisor_ctx.scan_time = scan_time;
+       advisor_ctx.cpu_time = cpu_time;
+
+       ksm_thread_pages_to_scan = pages;
+}
+
+static void advisor_stop_scan(void)
+{
+       if (ksm_advisor == KSM_ADVISOR_SCAN_TIME)
+               scan_time_advisor();
+}
+
 #ifdef CONFIG_NUMA
 /* Zeroed when merging across nodes is not allowed */
 static unsigned int ksm_merge_across_nodes = 1;
@@ -2401,6 +2551,7 @@ static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page)
 
        mm_slot = ksm_scan.mm_slot;
        if (mm_slot == &ksm_mm_head) {
+               advisor_start_scan();
                trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items);
 
                /*
@@ -2558,6 +2709,8 @@ no_vmas:
        if (mm_slot != &ksm_mm_head)
                goto next_mm;
 
+       advisor_stop_scan();
+
        trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items);
        ksm_scan.seqnr++;
        return NULL;
@@ -3246,6 +3399,9 @@ static ssize_t pages_to_scan_store(struct kobject *kobj,
        unsigned int nr_pages;
        int err;
 
+       if (ksm_advisor != KSM_ADVISOR_NONE)
+               return -EINVAL;
+
        err = kstrtouint(buf, 10, &nr_pages);
        if (err)
                return -EINVAL;