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 29a4a640b529ac77c79564dd9a7b3c76041c9301..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
@@ -1005,13 +1015,6 @@ Clusters:
       # with the cancelled container.
       MaxRetryAttempts: 3
 
-      # The maximum number of compute nodes that can be in use simultaneously
-      # If this limit is reduced, any existing nodes with slot number >= new limit
-      # will not be counted against the new limit. In other words, the new limit
-      # won't be strictly enforced until those nodes with higher slot numbers
-      # go down.
-      MaxComputeVMs: 64
-
       # Schedule all child containers on preemptible instances (e.g. AWS
       # Spot Instances) even if not requested by the submitter.
       #
@@ -1327,6 +1330,30 @@ Clusters:
         # providers too, if desired.
         MaxConcurrentInstanceCreateOps: 1
 
+        # The maximum number of instances to run at a time, or 0 for
+        # unlimited.
+        #
+        # If more instances than this are already running and busy
+        # when the dispatcher starts up, the running containers will
+        # be allowed to finish before the excess instances are shut
+        # 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
@@ -1779,9 +1806,11 @@ Clusters:
       # This feature is disabled when set to zero.
       IdleTimeout: 0s
 
-      # URL to a file that is a fragment of text or HTML which should
-      # be rendered in Workbench as a banner.
-      BannerURL: ""
+      # UUID of a collection.  This collection should be shared with
+      # all users.  Workbench will look for a file "banner.html" in
+      # this collection and display its contents (should be
+      # HTML-formatted text) when users first log in to Workbench.
+      BannerUUID: ""
 
       # Workbench welcome screen, this is HTML text that will be
       # incorporated directly onto the page.