CORE-2611 Changed prerendering to vCPU scaling, give each task at least 2 vCPU (1 per thread)

script/prerender/cfn.yml

@@ -42,11 +42,11 @@ Parameters:
     AllowedPattern: ^\d{4}-\d\d-\d\dT\d\d:\d\d:\d\dZ$
     ConstraintDescription: Must be a timestamp in the YYYY-MM-DDTHH:MM:SSZ format
 
-  TotalMemoryMiB:
-    Description: Total fleet memory capacity in MiB (must equal desired task count * 4096)
+  TotalVCpuCount:
+    Description: Total fleet vCPU capacity (must equal desired task count * 2)
     Type: Number
-    MinValue: 4096
-    ConstraintDescription: Must equal desired task count * 4096
+    MinValue: 2
+    ConstraintDescription: Must equal desired task count * 2
 
 Resources:
 
@@ -81,6 +81,7 @@ Resources:
       ContainerDefinitions:
         - Command:
             - prerender:work
+          Cpu: 2048
           EntryPoint:
             - yarn
           Environment:
@@ -109,7 +110,7 @@ Resources:
               awslogs-region: !Ref AWS::Region
               awslogs-stream-prefix: workers
               mode: non-blocking
-          MemoryReservation: 4096
+          MemoryReservation: 3686
           Name: !Sub ${AWS::StackName}-task-definition
       Family: Prerendering
       NetworkMode: host
@@ -123,6 +124,13 @@ Resources:
       EnableECSManagedTags: true
       EnableExecuteCommand: true
       LaunchType: EC2
+      PlacementStrategies:
+        - Field: attribute:ecs.availability-zone
+          Type: spread
+        - Field: cpu
+          Type: binpack
+        - Field: memory
+          Type: binpack
       PropagateTags: SERVICE
       ServiceName: Prerendering
       TaskDefinition: !Ref TaskDefinition
@@ -198,9 +206,7 @@ Resources:
                 MaxSpotPriceAsPercentageOfOptimalOnDemandPrice: 101
                 MemoryGiBPerVCpu:
                   Min: 2
-                  Max: 8
                 MemoryMiB:
-                  Max: !Ref TotalMemoryMiB
                   Min: 4096
                 VCpuCount:
                   Min: 2
@@ -223,8 +229,8 @@ Resources:
               Value: !Sub ${AWS::StackName}-fleet
       TargetCapacitySpecification:
         DefaultTargetCapacityType: spot
-        TargetCapacityUnitType: memory-mib
-        TotalTargetCapacity: !Ref TotalMemoryMiB
+        TargetCapacityUnitType: vcpu
+        TotalTargetCapacity: !Ref TotalVCpuCount
       TerminateInstancesWithExpiration: true
       ValidUntil: !Ref ValidUntil
 

script/prerender/fleet.ts

@@ -60,10 +60,10 @@ import { RELEASE_ID, WORK_REGION, BUCKET_NAME, BUCKET_REGION, PUBLIC_URL } from
 const MAX_CONCURRENT_BOOKS = 5;
 
 // Number of concurrent prerender tasks to run
-const DESIRED_TASK_COUNT = 32;
+const DESIRED_TASK_COUNT = 16;
 
-// Total fleet memory capacity in MiB (must equal DESIRED_TASK_COUNT * 4096)
-const TOTAL_MEMORY_MIB = DESIRED_TASK_COUNT * 4096;
+// Total fleet vCPU capacity (must equal DESIRED_TASK_COUNT * 2)
+const TOTAL_VCPU_COUNT = DESIRED_TASK_COUNT * 2;
 
 // Retry EPROTO errors in requests this many times
 const MAX_ATTEMPTS = 5;
@@ -200,8 +200,8 @@ async function createWorkersStack() {
         ParameterValue: RELEASE_ID,
       },
       {
-        ParameterKey: 'TotalMemoryMiB',
-        ParameterValue: TOTAL_MEMORY_MIB.toString(),
+        ParameterKey: 'TotalVCpuCount',
+        ParameterValue: TOTAL_VCPU_COUNT.toString(),
       },
       {
         ParameterKey: 'ValidUntil',

script/prerender/work.ts

@@ -21,8 +21,7 @@ import './logUnhandledRejectionsAndExit';
 // The timeout must be long enough to render the slowest page, otherwise builds will never finish
 const MAX_HEARTBEATS = 20;
 
-// Since SchedulingStrategy switched from DAEMON to REPLICA, MemoryReservation controls how many
-// tasks each instance gets; at a 2 GiB/vCPU floor, a 4096 MiB task can get up to 2 vCPUs
+// Each task reserves 2 vCPUs, one per thread
 const WORKER_THREAD_COUNT = 2;
 
 console.log(`Bucket: ${process.env.BUCKET_NAME} (${process.env.BUCKET_REGION})`);