planner: introduce SPOT VMs policy

.env

@@ -1,4 +1,4 @@
-FAABRIC_VERSION=0.18.0
-FAABRIC_CLI_IMAGE=faasm.azurecr.io/faabric:0.18.0
+FAABRIC_VERSION=0.19.0
+FAABRIC_CLI_IMAGE=faasm.azurecr.io/faabric:0.19.0
 COMPOSE_PROJECT_NAME=faabric-dev
 CONAN_CACHE_MOUNT_SOURCE=./conan-cache/

.github/workflows/tests.yml

@@ -20,7 +20,7 @@ jobs:
     if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
     env:
       DEPLOYMENT_TYPE: gha-ci
     steps:
@@ -34,7 +34,7 @@ jobs:
     if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
     steps:
       - name: "Check out code"
         uses: actions/checkout@v4
@@ -45,7 +45,7 @@ jobs:
     if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
     steps:
       - name: "Check out code"
         uses: actions/checkout@v4
@@ -65,7 +65,7 @@ jobs:
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
       options: --privileged
     services:
       redis:
@@ -104,7 +104,7 @@ jobs:
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
       options: --privileged
     services:
       redis:
@@ -156,7 +156,7 @@ jobs:
       REDIS_QUEUE_HOST: redis
       REDIS_STATE_HOST: redis
     container:
-      image: faasm.azurecr.io/faabric:0.18.0
+      image: faasm.azurecr.io/faabric:0.19.0
     services:
       redis:
         image: redis

VERSION

@@ -1 +1 @@
-0.18.0
+0.19.0

include/faabric/batch-scheduler/BatchScheduler.h

@@ -12,6 +12,11 @@
 #define NOT_ENOUGH_SLOTS_DECISION                                              \
     faabric::batch_scheduler::SchedulingDecision(NOT_ENOUGH_SLOTS,             \
                                                  NOT_ENOUGH_SLOTS)
+#define MUST_FREEZE -97
+#define MUST_FREEZE_DECISION                                                   \
+    faabric::batch_scheduler::SchedulingDecision(MUST_FREEZE, MUST_FREEZE)
+
+#define MUST_EVICT_IP "E.VI.CT.ME"
 
 namespace faabric::batch_scheduler {
 

include/faabric/batch-scheduler/SpotScheduler.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include <faabric/batch-scheduler/BatchScheduler.h>
+#include <faabric/batch-scheduler/SchedulingDecision.h>
+#include <faabric/util/batch.h>
+
+namespace faabric::batch_scheduler {
+
+// This batch scheduler behaves in the same way than BinPack for NEW and
+// SCALE_CHANGE requests, but for DIST_CHANGE it considers if any of the
+// hosts in the Host Map have been tainted with the eviction mark. In which
+// case it first tries to migrate them to other running hosts and, if not
+// enough hosts are available, freezes the messages.
+class SpotScheduler final : public BatchScheduler
+{
+  public:
+    std::shared_ptr<SchedulingDecision> makeSchedulingDecision(
+      HostMap& hostMap,
+      const InFlightReqs& inFlightReqs,
+      std::shared_ptr<faabric::BatchExecuteRequest> req) override;
+
+  private:
+    bool isFirstDecisionBetter(
+      std::shared_ptr<SchedulingDecision> decisionA,
+      std::shared_ptr<SchedulingDecision> decisionB) override;
+
+    std::vector<Host> getSortedHosts(
+      HostMap& hostMap,
+      const InFlightReqs& inFlightReqs,
+      std::shared_ptr<faabric::BatchExecuteRequest> req,
+      const DecisionType& decisionType) override;
+};
+}

include/faabric/planner/Planner.h

@@ -33,6 +33,8 @@ class Planner
 
     void printConfig() const;
 
+    std::string getPolicy();
+
     void setPolicy(const std::string& newPolicy);
 
     // ----------
@@ -87,10 +89,21 @@ class Planner
     // the planner was last reset
     int getNumMigrations();
 
+    // Helper method to get the next host that will be evicted
+    std::set<std::string> getNextEvictedHostIps();
+
+    std::map<int32_t, std::shared_ptr<BatchExecuteRequest>> getEvictedReqs();
+
     // Main entrypoint to request the execution of batches
     std::shared_ptr<faabric::batch_scheduler::SchedulingDecision> callBatch(
       std::shared_ptr<BatchExecuteRequest> req);
 
+    // ----------
+    // API exclusive to SPOT policy mode
+    // ----------
+
+    void setNextEvictedVm(const std::set<std::string>& vmIp);
+
   private:
     // There's a singleton instance of the planner running, but it must allow
     // concurrent requests

include/faabric/planner/PlannerState.h

@@ -12,6 +12,10 @@ namespace faabric::planner {
  */
 struct PlannerState
 {
+    // Policy to operate the planner in. Mostly determins the batch scheduler
+    // behaviour, but also the planner's in some cases
+    std::string policy;
+
     // Accounting of the hosts that are registered in the system and responsive
     // We deliberately use the host's IP as unique key, but assign a unique host
     // id for redundancy
@@ -36,5 +40,19 @@ struct PlannerState
 
     // Helper coutner of the total number of migrations
     std::atomic<int> numMigrations = 0;
+
+    // -----
+    // Data structures used only under the SPOT policy
+    // -----
+
+    // Map containing the BER that have been evicted due to a SPOT VM eviction.
+    // All messages in the VM have been checkpointed, are in the snapshot
+    // registry in the planner, and are ready to be scheduled when capacity
+    // appears
+    std::map<int, std::shared_ptr<BatchExecuteRequest>> evictedRequests;
+
+    // This variable simulates the values we would get from a cloud provider's
+    // API indicating the (set of) VM to be evicted next
+    std::set<std::string> nextEvictedHostIps;
 };
 }

include/faabric/util/func.h

@@ -6,9 +6,18 @@
 #include <vector>
 
 #define MIGRATED_FUNCTION_RETURN_VALUE -99
+#define FROZEN_FUNCTION_RETURN_VALUE -98
 
 namespace faabric::util {
 
+class FunctionFrozenException : public faabric::util::FaabricException
+{
+  public:
+    explicit FunctionFrozenException(std::string message)
+      : FaabricException(std::move(message))
+    {}
+};
+
 class FunctionMigratedException : public faabric::util::FaabricException
 {
   public:

src/batch-scheduler/BatchScheduler.cpp

@@ -1,6 +1,7 @@
 #include <faabric/batch-scheduler/BatchScheduler.h>
 #include <faabric/batch-scheduler/BinPackScheduler.h>
 #include <faabric/batch-scheduler/CompactScheduler.h>
+#include <faabric/batch-scheduler/SpotScheduler.h>
 #include <faabric/util/config.h>
 #include <faabric/util/logging.h>
 
@@ -23,6 +24,8 @@ std::shared_ptr<BatchScheduler> getBatchScheduler()
         batchScheduler = std::make_shared<BinPackScheduler>();
     } else if (mode == "compact") {
         batchScheduler = std::make_shared<CompactScheduler>();
+    } else if (mode == "spot") {
+        batchScheduler = std::make_shared<SpotScheduler>();
     } else {
         SPDLOG_ERROR("Unrecognised batch scheduler mode: {}", mode);
         throw std::runtime_error("Unrecognised batch scheduler mode");

src/batch-scheduler/CMakeLists.txt

@@ -7,6 +7,7 @@ faabric_lib(batch_scheduler
     BatchScheduler.cpp
     BinPackScheduler.cpp
     CompactScheduler.cpp
+    SpotScheduler.cpp
 )
 
 target_link_libraries(batch_scheduler PRIVATE

src/batch-scheduler/CompactScheduler.cpp

@@ -98,7 +98,7 @@ bool CompactScheduler::isFirstDecisionBetter(
     throw std::runtime_error("Method not supported for COMPACT scheduler");
 }
 
-HostMap deepCopyHostMap(const HostMap& hostMap)
+static HostMap deepCopyHostMap(const HostMap& hostMap)
 {
     HostMap newHostMap;
 
@@ -173,9 +173,9 @@ bool CompactScheduler::isFirstDecisionBetter(
 
 // Filter-out from the host map all nodes that are executing requests from a
 // different user
-void filterHosts(HostMap& hostMap,
-                 const InFlightReqs& inFlightReqs,
-                 std::shared_ptr<faabric::BatchExecuteRequest> req)
+static void filterHosts(HostMap& hostMap,
+                        const InFlightReqs& inFlightReqs,
+                        std::shared_ptr<faabric::BatchExecuteRequest> req)
 {
     // We temporarily use the request subtype field to attach a user id for our
     // multi-tenant simulations