20680: Set concurrency to num_cpus * 2 with a minimum of 8

[arvados.git] / tools / salt-install / config_examples / multi_host / aws / pillars / nginx_passenger.sls

diff --git a/tools/salt-install/config_examples/multi_host/aws/pillars/nginx_passenger.sls b/tools/salt-install/config_examples/multi_host/aws/pillars/nginx_passenger.sls
index 7bbf9ae618a3c065e78af5ecfaad8d126c98eb62..4c0aea25fe7ada8c6b9cf0f0853df8f56fdd2f52 100644 (file)
--- a/tools/salt-install/config_examples/multi_host/aws/pillars/nginx_passenger.sls
+++ b/tools/salt-install/config_examples/multi_host/aws/pillars/nginx_passenger.sls
@@ -12,6 +12,9 @@
 {%- set passenger_ruby = '/usr/local/rvm/wrappers/default/ruby'
                            if grains.osfinger in ('CentOS Linux-7', 'Ubuntu-18.04', 'Debian-10') else
                          '/usr/bin/ruby' %}
+{%- set _workers = ("__CONTROLLER_MAX_WORKERS__" or grains['num_cpus']*2)|int %}
+{%- set max_workers = [_workers, 8]|max %}
+{%- set max_reqs = ("__CONTROLLER_MAX_QUEUED_REQUESTS__" or 128)|int %}
 
 ### NGINX
 nginx:
@@ -21,13 +24,15 @@ nginx:
   ### PASSENGER
   passenger:
     passenger_ruby: {{ passenger_ruby }}
-    passenger_max_pool_size: {{ "__CONTROLLER_NGINX_WORKERS__" or grains['num_cpus'] }}
-    {%- set max_reqs = "__CONTROLLER_MAX_CONCURRENT_REQUESTS__" %}
-    {%- if max_reqs != "" and max_reqs is number %}
-    # Default is 100 -- Configuring this a bit higher than API.MaxConcurrentRequests
-    # to be able to handle /metrics requests even on heavy load situations.
-    passenger_max_request_queue_size: {{ (max_reqs * 1.1)|round|int }}
-    {%- endif %}
+    passenger_max_pool_size: {{ max_workers }}
+
+    # Make the passenger queue small (twice the concurrency, so
+    # there's at most one pending request for each busy worker)
+    # because controller reorders requests based on priority, and
+    # won't send more than API.MaxConcurrentRequests to passenger
+    # (which is max_workers * 2), so things that are moved to the head
+    # of the line get processed quickly.
+    passenger_max_request_queue_size: {{ max_workers * 2 + 1 }}
 
   ### SERVER
   server:
@@ -43,10 +48,15 @@ nginx:
       # include: 'modules-enabled/*.conf'
       load_module: {{ passenger_mod }}
       {% endif %}
-      worker_processes: {{ "__CONTROLLER_NGINX_WORKERS__" or grains['num_cpus'] }}
-      worker_rlimit_nofile: 4096
+      worker_processes: {{ max_workers }}
+
+      # each request is up to 3 connections (1 with client, 1 proxy to
+      # controller, then potentially 1 from controller back to
+      # passenger).  Each connection consumes a file descriptor.
+      # That's how we get these calculations
+      worker_rlimit_nofile: {{ max_reqs * 3 + 1 }}
       events:
-        worker_connections: 1024
+        worker_connections: {{ max_reqs * 3 + 1 }}
 
   ### SNIPPETS
   snippets: