From 31b8aa3c83f2bb1cd448c40677a1eb8266dad2ce Mon Sep 17 00:00:00 2001
From: Boris Kolpackov <boris@codesynthesis.com>
Date: Mon, 15 May 2023 11:58:53 +0200
Subject: Add CPU boosting for higher 4th priority levels

---
 bbot/agent/agent.cli   |  6 +++++-
 bbot/agent/agent.cxx   | 16 +++++++++++++---
 bbot/agent/machine.cxx | 13 ++++++++-----
 bbot/agent/machine.hxx |  3 ++-
 4 files changed, 28 insertions(+), 10 deletions(-)

diff --git a/bbot/agent/agent.cli b/bbot/agent/agent.cli
index 1046685..0a9147c 100644
--- a/bbot/agent/agent.cli
+++ b/bbot/agent/agent.cli
@@ -60,7 +60,11 @@ namespace bbot
     number (\cb{--cpu}) is boosted to the full number of available hardware
     threads (or, to view it another way, the fourth priority level has 20
     possible values, not 10, with the first 0-9 being without the boost while
-    the last 10-19 being with the boost).
+    the last 10-19 being with the boost). Note that this boosting semantics
+    may not be accurate if the agent is executed with CPU affinity. Also note
+    that there is no corresponding RAM boosting and it's possible that in some
+    configurations the amount of RAM will be insufficient for the boosted CPU
+    count.
 
     Note that the priority levels are hierarchical in a sense that within a
     given higher level URLs can be further prioritized using the lower
diff --git a/bbot/agent/agent.cxx b/bbot/agent/agent.cxx
index 67cf57f..9aba216 100644
--- a/bbot/agent/agent.cxx
+++ b/bbot/agent/agent.cxx
@@ -23,6 +23,7 @@
 #include <map>
 #include <atomic>
 #include <chrono>
+#include <thread>       // thread::hardware_concurrency()
 #include <random>
 #include <iomanip>      // setw()
 #include <iostream>
@@ -1250,7 +1251,8 @@ perform_task (toolchain_lock tl, // Note: assumes ownership.
               machine_lock& ml,
               const dir_path& md,
               const bootstrapped_machine_manifest& mm,
-              const task_manifest& tm)
+              const task_manifest& tm,
+              optional<size_t> boost_cpus)
 try
 {
   tracer trace ("perform_task", md.string ().c_str ());
@@ -1399,7 +1401,8 @@ try
                          mm.machine.mac,
                          ops.bridge (),
                          tftpd.port (),
-                         tm.interactive.has_value ()));
+                         tm.interactive.has_value (),
+                         boost_cpus));
 
         auto mg (
           make_exception_guard (
@@ -2525,8 +2528,15 @@ try
         }
       }
 
+      // Check if we need to boost the number of CPUs to the full hardware
+      // concurrency.
+      //
+      optional<size_t> bcpus;
+      if (prio >= 10000)
+        bcpus = std::thread::hardware_concurrency ();
+
       pm->lock.perform_task (tl, prio);
-      r = perform_task (move (tl), pm->lock, pm->path, pm->manifest, t);
+      r = perform_task (move (tl), pm->lock, pm->path, pm->manifest, t, bcpus);
     }
     catch (const interrupt&)
     {
diff --git a/bbot/agent/machine.cxx b/bbot/agent/machine.cxx
index 8308989..16f0525 100644
--- a/bbot/agent/machine.cxx
+++ b/bbot/agent/machine.cxx
@@ -172,7 +172,8 @@ namespace bbot
                  const optional<string>& mac,
                  const string& br_iface,
                  uint16_t tftp_port,
-                 bool pub_vnc);
+                 bool pub_vnc,
+                 optional<size_t> boost_cpus = nullopt);
 
     virtual bool
     shutdown (size_t& seconds) override;
@@ -216,7 +217,8 @@ namespace bbot
                const optional<string>& omac,
                const string& br,
                uint16_t port,
-               bool pub_vnc)
+               bool pub_vnc,
+               optional<size_t> bcpus)
       : machine (mm.mac ? *mm.mac : // Fixed mac from machine manifest.
                  omac   ? *omac   : // Generated mac from previous bootstrap.
                  generate_mac ()),
@@ -249,7 +251,7 @@ namespace bbot
     // Note that for best results you may want to adjust (e.g., by over-
     // committing) the number of CPUs to be power of 2.
     //
-    size_t cpus (ops.cpu ()), cores (cpus);
+    size_t cpus (bcpus ? *bcpus : ops.cpu ()), cores (cpus);
 
     size_t sockets (cores >= 16 && cores % 4 == 0 ? 2 :
                     cores >= 64 && cores % 8 == 0 ? 4 : 1);
@@ -672,13 +674,14 @@ namespace bbot
                  const optional<string>& mac,
                  const string& br_iface,
                  uint16_t tftp_port,
-                 bool pub_vnc)
+                 bool pub_vnc,
+                 optional<size_t> bcpus)
   {
     switch (mm.type)
     {
     case machine_type::kvm:
       return make_unique<kvm_machine> (
-        md, mm, mac, br_iface, tftp_port, pub_vnc);
+        md, mm, mac, br_iface, tftp_port, pub_vnc, bcpus);
 
     case machine_type::nspawn:
       assert (false); //@@ TODO
diff --git a/bbot/agent/machine.hxx b/bbot/agent/machine.hxx
index adace1f..0bb74b9 100644
--- a/bbot/agent/machine.hxx
+++ b/bbot/agent/machine.hxx
@@ -84,7 +84,8 @@ namespace bbot
                  const optional<string>& mac,
                  const string& br_iface,
                  uint16_t tftp_port,
-                 bool pub_vnc);
+                 bool pub_vnc,
+                 optional<size_t> boost_cpus = nullopt);
 
   // Return the machine's public or private VNC session endpoint in the
   // '<ip>:<port>' form.
-- 
cgit v1.1