From ef88a00d899009283143dfb64c3f9629c0b6e31d Mon Sep 17 00:00:00 2001
From: Joseph Coffland <joseph@cauldrondevelopment.com>
Date: Thu, 9 Mar 2023 18:36:14 +0200
Subject: [PATCH] Fix CPU allocation when there are more GPUs than CPUs. #129,
 Don't reserve a CPU for each disabled GPUs.

---
 CHANGELOG.md             |  4 +++-
 src/fah/client/App.cpp   |  2 +-
 src/fah/client/Units.cpp | 14 ++++----------
 3 files changed, 8 insertions(+), 12 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index ae47260..355e361 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,7 +4,9 @@ Folding@home Client Changelog
 ## v8.1.16
  - Fix core download retry logic.
  - Only add client executable directory to lib path on Windows.
- - Retry WU if core crashes.  #127
+ - Retry WU if core crashes. #127
+ - Fix CPU allocation when there are more GPUs than CPUs. #129
+ - Don't reserve a CPU for each disabled GPUs.
 
 ## v8.1.15
  - Fix CUDA/OpenCL driver mixup from v8.1.14.
diff --git a/src/fah/client/App.cpp b/src/fah/client/App.cpp
index 2324235..8992f04 100644
--- a/src/fah/client/App.cpp
+++ b/src/fah/client/App.cpp
@@ -301,7 +301,7 @@ void App::updateResources() {
     uint32_t cpus = min(config.getCPUs(), availableCPUs);
 
     info->insert("cpus", cpus + remainingCPUs);
-    availableCPUs -= cpus;
+    availableCPUs -= min(availableCPUs, cpus);
     if (cpus < config.getCPUs()) config.insert("cpus", cpus);
 
     JSON::ValuePtr groupGPUs = new JSON::Dict;
diff --git a/src/fah/client/Units.cpp b/src/fah/client/Units.cpp
index 251c1aa..90f2a82 100644
--- a/src/fah/client/Units.cpp
+++ b/src/fah/client/Units.cpp
@@ -200,22 +200,18 @@ void Units::update() {
     if (unitGPUs.empty()) continue;
 
     uint32_t minCPUs = unit.getMinCPUs();
-    bool     runable = minCPUs <= remainingCPUs;
+    bool     runable = minCPUs <= remainingCPUs || minCPUs < 2;
 
     std::set<string> gpusWithWU = remainingGPUs;
     for (auto id: unitGPUs) runable |= gpusWithWU.erase(id);
 
     if (runable) {
       remainingGPUs = gpusWithWU;
-      remainingCPUs -= minCPUs; // Initially allocate only minimum CPUs
+      remainingCPUs -= min(remainingCPUs, minCPUs); // Allocate minimum CPUs
       enabledWUs.insert(i);
     }
   }
 
-  // Reserve one CPU for any unused GPUs
-  uint32_t reservedCPUs = min(remainingCPUs, (uint32_t)remainingGPUs.size());
-  remainingCPUs -= reservedCPUs;
-
   // Allocate extra CPUs to enabled GPU WUs
   for (unsigned i = 0; i < size(); i++) {
     if (!enabledWUs.count(i)) continue; // GPU WUs that were enabled above
@@ -226,7 +222,8 @@ void Units::update() {
     uint32_t cpus    = min(maxCPUs, remainingCPUs + minCPUs);
 
     unit.setCPUs(cpus);
-    remainingCPUs -= cpus - minCPUs; // Minimum CPUs subtracted above
+    // minCPUs was subtracted above or remainingCPUs is already zero
+    remainingCPUs -= min(remainingCPUs, cpus - minCPUs);
   }
 
   // Allocate remaining CPUs to existing CPU WUs
@@ -243,9 +240,6 @@ void Units::update() {
     enabledWUs.insert(i);
   }
 
-  // Restore reserved CPUs
-  remainingCPUs += reservedCPUs;
-
   // Start and stop WUs
   for (unsigned i = 0; i < size(); i++)
     getUnit(i).setPause(getUnit(i).atRunState() && !enabledWUs.count(i));