Add config flags to allow for cache auto-tuning

changelog.d/12701.feature

@@ -0,0 +1 @@
+Add a config options to allow for auto-tuning of caches.

docs/sample_config.yaml

@@ -784,6 +784,24 @@ caches:
   #
   #cache_entry_ttl: 30m
 
+  # This flag enables cache autotuning, and is further specified by the sub-options `max_cache_memory_usage`,
+  # `target_cache_memory_usage`, `min_cache_ttl`. These flags work in conjunction with each other to maintain
+  # a balance between cache memory usage and cache entry availability. You must be using jemalloc to utilize
+  # this option, and all three of the options must be specified for this feature to work.
+  #cache_autotuning:
+    # This flag sets a ceiling on much memory the cache can use before caches begin to be continuously evicted.
+    # They will continue to be evicted until the memory usage drops below the `target_memory_usage`, set in
+    # the flag below, or until the `min_cache_ttl` is hit.
+    #max_cache_memory_usage: 1024M
+
+    # This flag sets a rough target for the desired memory usage of the caches.
+    #target_cache_memory_usage: 758M
+
+    # 'min_cache_ttl` sets a limit under which newer cache entries are not evicted and is only applied when
+    # caches are actively being evicted/`max_cache_memory_usage` has been exceeded. This is to protect hot caches
+    # from being emptied while Synapse is evicting due to memory.
+    #min_cache_ttl: 5m
+
   # Controls how long the results of a /sync request are cached for after
   # a successful response is returned. A higher duration can help clients with
   # intermittent connections, at the cost of higher memory usage.

docs/usage/configuration/config_documentation.md

@@ -1119,16 +1119,29 @@ Caching can be configured through the following sub-options:
   with intermittent connections, at the cost of higher memory usage.
   By default, this is zero, which means that sync responses are not cached
   at all.
-
+* `cache_autotuning` and its sub-options `max_cache_memory_usage`, `target_cache_memory_usage`, and
+   `min_cache_ttl` work in conjunction with each other to maintain a balance between cache memory 
+   usage and cache entry availability. You must be using [jemalloc](https://github.com/matrix-org/synapse#help-synapse-is-slow-and-eats-all-my-ramcpu) 
+   to utilize this option, and all three of the options must be specified for this feature to work.
+     * `max_cache_memory_usage` sets a ceiling on how much memory the cache can use before caches begin to be continuously evicted.
+        They will continue to be evicted until the memory usage drops below the `target_memory_usage`, set in
+        the flag below, or until the `min_cache_ttl` is hit.
+     * `target_memory_usage` sets a rough target for the desired memory usage of the caches.
+     * `min_cache_ttl` sets a limit under which newer cache entries are not evicted and is only applied when
+        caches are actively being evicted/`max_cache_memory_usage` has been exceeded. This is to protect hot caches
+        from being emptied while Synapse is evicting due to memory.
 
 Example configuration:
 ```yaml
 caches:
   global_factor: 1.0
   per_cache_factors:
     get_users_who_share_room_with_user: 2.0
-  expire_caches: false
   sync_response_cache_duration: 2m
+  cache_autotuning:
+    max_cache_memory_usage: 1024M
+    target_cache_memory_usage: 758M
+    min_cache_ttl: 5m
 ```
 
 ### Reloading cache factors

synapse/config/cache.py

@@ -176,6 +176,24 @@ def generate_config_section(self, **kwargs: Any) -> str:
           #
           #cache_entry_ttl: 30m
 
+          # This flag enables cache autotuning, and is further specified by the sub-options `max_cache_memory_usage`,
+          # `target_cache_memory_usage`, `min_cache_ttl`. These flags work in conjunction with each other to maintain
+          # a balance between cache memory usage and cache entry availability. You must be using jemalloc to utilize
+          # this option, and all three of the options must be specified for this feature to work.
+          #cache_autotuning:
+            # This flag sets a ceiling on much memory the cache can use before caches begin to be continuously evicted.
+            # They will continue to be evicted until the memory usage drops below the `target_memory_usage`, set in
+            # the flag below, or until the `min_cache_ttl` is hit.
+            #max_cache_memory_usage: 1024M
+
+            # This flag sets a rough target for the desired memory usage of the caches.
+            #target_cache_memory_usage: 758M
+
+            # 'min_cache_ttl` sets a limit under which newer cache entries are not evicted and is only applied when
+            # caches are actively being evicted/`max_cache_memory_usage` has been exceeded. This is to protect hot caches
+            # from being emptied while Synapse is evicting due to memory.
+            #min_cache_ttl: 5m
+
           # Controls how long the results of a /sync request are cached for after
           # a successful response is returned. A higher duration can help clients with
           # intermittent connections, at the cost of higher memory usage.
@@ -263,6 +281,21 @@ def read_config(self, config: JsonDict, **kwargs: Any) -> None:
             )
             self.expiry_time_msec = self.parse_duration(expiry_time)
 
+        self.cache_autotuning = cache_config.get("cache_autotuning")
+        if self.cache_autotuning:
+            max_memory_usage = self.cache_autotuning.get("max_cache_memory_usage")
+            self.cache_autotuning["max_cache_memory_usage"] = self.parse_size(
+                max_memory_usage
+            )
+
+            target_mem_size = self.cache_autotuning.get("target_cache_memory_usage")
+            self.cache_autotuning["target_cache_memory_usage"] = self.parse_size(
+                target_mem_size
+            )
+
+            min_cache_ttl = self.cache_autotuning.get("min_cache_ttl")
+            self.cache_autotuning["min_cache_ttl"] = self.parse_duration(min_cache_ttl)
+
         self.sync_response_cache_duration = self.parse_duration(
             cache_config.get("sync_response_cache_duration", 0)
         )

synapse/metrics/jemalloc.py

@@ -18,6 +18,7 @@
 import re
 from typing import Iterable, Optional, overload
 
+import attr
 from prometheus_client import REGISTRY, Metric
 from typing_extensions import Literal
 
@@ -27,52 +28,24 @@
 logger = logging.getLogger(__name__)
 
 
-def _setup_jemalloc_stats() -> None:
-    """Checks to see if jemalloc is loaded, and hooks up a collector to record
-    statistics exposed by jemalloc.
-    """
-
-    # Try to find the loaded jemalloc shared library, if any. We need to
-    # introspect into what is loaded, rather than loading whatever is on the
-    # path, as if we load a *different* jemalloc version things will seg fault.
-
-    # We look in `/proc/self/maps`, which only exists on linux.
-    if not os.path.exists("/proc/self/maps"):
-        logger.debug("Not looking for jemalloc as no /proc/self/maps exist")
-        return
-
-    # We're looking for a path at the end of the line that includes
-    # "libjemalloc".
-    regex = re.compile(r"/\S+/libjemalloc.*$")
-
-    jemalloc_path = None
-    with open("/proc/self/maps") as f:
-        for line in f:
-            match = regex.search(line.strip())
-            if match:
-                jemalloc_path = match.group()
-
-    if not jemalloc_path:
-        # No loaded jemalloc was found.
-        logger.debug("jemalloc not found")
-        return
-
-    logger.debug("Found jemalloc at %s", jemalloc_path)
-
-    jemalloc = ctypes.CDLL(jemalloc_path)
+@attr.s(slots=True, frozen=True, auto_attribs=True)
+class JemallocStats:
+    jemalloc: ctypes.CDLL
 
     @overload
     def _mallctl(
-        name: str, read: Literal[True] = True, write: Optional[int] = None
+        self, name: str, read: Literal[True] = True, write: Optional[int] = None
     ) -> int:
         ...
 
     @overload
-    def _mallctl(name: str, read: Literal[False], write: Optional[int] = None) -> None:
+    def _mallctl(
+        self, name: str, read: Literal[False], write: Optional[int] = None
+    ) -> None:
         ...
 
     def _mallctl(
-        name: str, read: bool = True, write: Optional[int] = None
+        self, name: str, read: bool = True, write: Optional[int] = None
     ) -> Optional[int]:
         """Wrapper around `mallctl` for reading and writing integers to
         jemalloc.
@@ -120,7 +93,7 @@ def _mallctl(
         # Where oldp/oldlenp is a buffer where the old value will be written to
         # (if not null), and newp/newlen is the buffer with the new value to set
         # (if not null). Note that they're all references *except* newlen.
-        result = jemalloc.mallctl(
+        result = self.jemalloc.mallctl(
             name.encode("ascii"),
             input_var_ref,
             input_len_ref,
@@ -136,21 +109,80 @@ def _mallctl(
 
         return input_var.value
 
-    def _jemalloc_refresh_stats() -> None:
+    def refresh_stats(self) -> None:
         """Request that jemalloc updates its internal statistics. This needs to
         be called before querying for stats, otherwise it will return stale
         values.
         """
         try:
-            _mallctl("epoch", read=False, write=1)
+            self._mallctl("epoch", read=False, write=1)
         except Exception as e:
             logger.warning("Failed to reload jemalloc stats: %s", e)
 
+    def get_stat(self, name: str) -> int:
+        """Request the stat of the given name at the time of the last
+        `refresh_stats` call. This may throw if we fail to read
+        the stat.
+        """
+        return self._mallctl(f"stats.{name}")
+
+
+_JEMALLOC_STATS: Optional[JemallocStats] = None
+
+
+def get_jemalloc_stats() -> Optional[JemallocStats]:
+    """Returns an interface to jemalloc, if it is being used.
+
+    Note that this will always return None until `setup_jemalloc_stats` has been
+    called.
+    """
+    return _JEMALLOC_STATS
+
+
+def _setup_jemalloc_stats() -> None:
+    """Checks to see if jemalloc is loaded, and hooks up a collector to record
+    statistics exposed by jemalloc.
+    """
+
+    global _JEMALLOC_STATS
+
+    # Try to find the loaded jemalloc shared library, if any. We need to
+    # introspect into what is loaded, rather than loading whatever is on the
+    # path, as if we load a *different* jemalloc version things will seg fault.
+
+    # We look in `/proc/self/maps`, which only exists on linux.
+    if not os.path.exists("/proc/self/maps"):
+        logger.debug("Not looking for jemalloc as no /proc/self/maps exist")
+        return
+
+    # We're looking for a path at the end of the line that includes
+    # "libjemalloc".
+    regex = re.compile(r"/\S+/libjemalloc.*$")
+
+    jemalloc_path = None
+    with open("/proc/self/maps") as f:
+        for line in f:
+            match = regex.search(line.strip())
+            if match:
+                jemalloc_path = match.group()
+
+    if not jemalloc_path:
+        # No loaded jemalloc was found.
+        logger.debug("jemalloc not found")
+        return
+
+    logger.debug("Found jemalloc at %s", jemalloc_path)
+
+    jemalloc_dll = ctypes.CDLL(jemalloc_path)
+
+    stats = JemallocStats(jemalloc_dll)
+    _JEMALLOC_STATS = stats
+
     class JemallocCollector(Collector):
         """Metrics for internal jemalloc stats."""
 
         def collect(self) -> Iterable[Metric]:
-            _jemalloc_refresh_stats()
+            stats.refresh_stats()
 
             g = GaugeMetricFamily(
                 "jemalloc_stats_app_memory_bytes",
@@ -184,7 +216,7 @@ def collect(self) -> Iterable[Metric]:
                 "metadata",
             ):
                 try:
-                    value = _mallctl(f"stats.{t}")
+                    value = stats.get_stat(t)
                 except Exception as e:
                     # There was an error fetching the value, skip.
                     logger.warning("Failed to read jemalloc stats.%s: %s", t, e)