20227: Change default BalacePeriod to 6 hours.

[arvados.git] / lib / config / config.default.yml

diff --git a/lib/config/config.default.yml b/lib/config/config.default.yml
index d1913563476547ecdf99f1d617c87e3751e25463..1919d7b704af0ce7c0d92fe4d0320229f84d0c51 100644 (file)
--- a/lib/config/config.default.yml
+++ b/lib/config/config.default.yml
@@ -225,7 +225,12 @@ Clusters:
 
       # Maximum number of concurrent requests to accept in a single
       # service process, or 0 for no limit.
-      MaxConcurrentRequests: 0
+      MaxConcurrentRequests: 64
+
+      # Fraction of MaxConcurrentRequests that can be "log create"
+      # messages at any given time.  This is to prevent logging
+      # updates from crowding out more important requests.
+      LogCreateRequestFraction: 0.50
 
       # Maximum number of 64MiB memory buffers per Keepstore server process, or
       # 0 for no limit. When this limit is reached, up to
@@ -529,7 +534,7 @@ Clusters:
       #
       # If SIGUSR1 is received during an idle period between operations,
       # the next operation will start immediately.
-      BalancePeriod: 10m
+      BalancePeriod: 6h
 
       # Limits the number of collections retrieved by keep-balance per
       # API transaction. If this is zero, page size is
@@ -992,6 +997,11 @@ Clusters:
       # disk cache size will use a disk cache, sized to the
       # container's RAM requirement (but with minimum 2 GiB and
       # maximum 32 GiB).
+      #
+      # Note: If you change this value, containers that used the previous
+      # default value will only be reused by container requests that
+      # explicitly specify the previous value in their keep_cache_ram
+      # runtime constraint.
       DefaultKeepCacheRAM: 0
 
       # Number of times a container can be unlocked before being
@@ -1329,6 +1339,21 @@ Clusters:
         # down.
         MaxInstances: 64
 
+        # Maximum fraction of CloudVMs.MaxInstances allowed to run
+        # "supervisor" containers at any given time. A supervisor is a
+        # container whose purpose is mainly to submit and manage other
+        # containers, such as arvados-cwl-runner workflow runner.
+        #
+        # If there is a hard limit on the amount of concurrent
+        # containers that the cluster can run, it is important to
+        # avoid crowding out the containers doing useful work with
+        # containers who just create more work.
+        #
+        # For example, with the default MaxInstances of 64, it will
+        # schedule at most floor(64*0.30) = 19 concurrent workflows,
+        # ensuring 45 slots are available for work.
+        SupervisorFraction: 0.30
+
         # Interval between cloud provider syncs/updates ("list all
         # instances").
         SyncInterval: 1m