Optimize thread_worker

Use a ring buffer to bulk pass tasks around, and also support creating a twin worker thread for every reactor thread. Before: ``` test iterations median mad min max allocs tasks inst thread_worker_test.1 3354 216.749us 47.062ns 216.590us 216.951us 5.000 4.000 1610844.2 thread_worker_test.10 460 2.088ms 870.335ns 2.087ms 2.089ms 70.000 58.000 25391981.4 thread_worker_test.100 61 16.465ms 1.733us 16.461ms 16.469ms 700.000 598.000 208106393.1 thread_worker_test.1000 50 19.925ms 6.187us 19.919ms 19.937ms 7007.000 5998.000 214776807.2 ``` After: ``` test iterations median mad min max allocs tasks inst thread_worker_test.1 1855 216.515us 77.552ns 216.330us 216.593us 4.000 4.000 1549639.0 thread_worker_test.10 413 2.085ms 394.971ns 2.084ms 2.086ms 40.000 40.000 24921287.6 thread_worker_test.100 113 8.518ms 3.891us 8.513ms 8.524ms 400.000 400.000 106121583.5 thread_worker_test.1000 93 10.424ms 605.968ns 10.423ms 10.426ms 4880.000 5744.000 127114775.7 ``` Signed-off-by: Tyler Rockwood <[email protected]>
redpanda-data · Aug 28, 2023 · 392564c · 392564c
1 parent ee5f900
commit 392564c
Show file tree

Hide file tree

Showing 10 changed files with 184 additions and 55 deletions.
diff --git a/src/v/kafka/server/server.cc b/src/v/kafka/server/server.cc
@@ -101,7 +101,7 @@ server::server(
   ss::sharded<cluster::tx_gateway_frontend>& tx_gateway_frontend,
   ss::sharded<cluster::tx_registry_frontend>& tx_registry_frontend,
   std::optional<qdc_monitor::config> qdc_config,
-  ssx::thread_worker& tw,
+  ssx::singleton_thread_worker& tw,
   const std::unique_ptr<pandaproxy::schema_registry::api>& sr) noexcept
   : net::server(cfg, klog)
   , _smp_group(smp)

diff --git a/src/v/kafka/server/server.h b/src/v/kafka/server/server.h
@@ -65,7 +65,7 @@ class server final
       ss::sharded<cluster::tx_gateway_frontend>&,
       ss::sharded<cluster::tx_registry_frontend>&,
       std::optional<qdc_monitor::config>,
-      ssx::thread_worker&,
+      ssx::singleton_thread_worker&,
       const std::unique_ptr<pandaproxy::schema_registry::api>&) noexcept;
 
     ~server() noexcept override = default;
@@ -161,7 +161,7 @@ class server final
 
     latency_probe& latency_probe() { return *_probe; }
 
-    ssx::thread_worker& thread_worker() { return _thread_worker; }
+    ssx::singleton_thread_worker& thread_worker() { return _thread_worker; }
 
     const std::unique_ptr<pandaproxy::schema_registry::api>& schema_registry() {
         return _schema_registry;
@@ -222,7 +222,7 @@ class server final
     ssx::metrics::metric_groups _metrics
       = ssx::metrics::metric_groups::make_internal();
     std::unique_ptr<class latency_probe> _probe;
-    ssx::thread_worker& _thread_worker;
+    ssx::singleton_thread_worker& _thread_worker;
     std::unique_ptr<replica_selector> _replica_selector;
     const std::unique_ptr<pandaproxy::schema_registry::api>& _schema_registry;
 };

diff --git a/src/v/redpanda/application.cc b/src/v/redpanda/application.cc
@@ -1059,7 +1059,9 @@ void application::wire_up_redpanda_services(
         resources::available_memory::local().register_metrics();
     }).get();
 
-    construct_single_service(thread_worker);
+    construct_single_service(
+      thread_worker,
+      ssx::thread_worker_config{.pin_to_shard_core = false, .name = "thread"});
 
     // cluster
     syschecks::systemd_message("Initializing connection cache").get();

diff --git a/src/v/redpanda/application.h b/src/v/redpanda/application.h
@@ -143,7 +143,7 @@ class application {
 
     std::unique_ptr<cluster::controller> controller;
 
-    std::unique_ptr<ssx::thread_worker> thread_worker;
+    std::unique_ptr<ssx::singleton_thread_worker> thread_worker;
 
     ss::sharded<kafka::server> _kafka_server;
 

diff --git a/src/v/security/gssapi_authenticator.cc b/src/v/security/gssapi_authenticator.cc
@@ -169,7 +169,7 @@ class gssapi_authenticator::impl {
 };
 
 gssapi_authenticator::gssapi_authenticator(
-  ssx::thread_worker& thread_worker,
+  ssx::singleton_thread_worker& thread_worker,
   std::vector<gssapi_rule> rules,
   ss::sstring principal,
   ss::sstring keytab)

diff --git a/src/v/security/gssapi_authenticator.h b/src/v/security/gssapi_authenticator.h
@@ -21,7 +21,7 @@ class gssapi_authenticator final : public sasl_mechanism {
     static constexpr const char* name = "GSSAPI";
 
     gssapi_authenticator(
-      ssx::thread_worker& thread_worker,
+      ssx::singleton_thread_worker& thread_worker,
       std::vector<gssapi_rule> rules,
       ss::sstring principal,
       ss::sstring keytab);
@@ -40,7 +40,7 @@ class gssapi_authenticator final : public sasl_mechanism {
     friend std::ostream&
     operator<<(std::ostream& os, gssapi_authenticator::state const s);
 
-    ssx::thread_worker& _worker;
+    ssx::singleton_thread_worker& _worker;
     security::acl_principal _principal;
     state _state{state::init};
     class impl;

diff --git a/src/v/ssx/fwd.h b/src/v/ssx/fwd.h
@@ -13,6 +13,6 @@
 
 namespace ssx {
 
-class thread_worker;
+class singleton_thread_worker;
 
 } // namespace ssx
diff --git a/src/v/ssx/tests/thread_worker.cc b/src/v/ssx/tests/thread_worker.cc
@@ -34,7 +34,7 @@ template<size_t tries, size_t stop_at>
 auto thread_worker_test() {
     BOOST_REQUIRE_GT(ss::smp::count, 1);
 
-    auto w = ssx::thread_worker{};
+    auto w = ssx::singleton_thread_worker{{}};
     w.start().get();
 
     std::vector<std::vector<ss::future<move_only>>> all_results(ss::smp::count);

diff --git a/src/v/ssx/tests/thread_worker_bench.cc b/src/v/ssx/tests/thread_worker_bench.cc
@@ -13,7 +13,7 @@
 #include <seastar/testing/perf_tests.hh>
 
 ss::future<> run_test(size_t data_size) {
-    auto w = ssx::thread_worker{};
+    auto w = ssx::singleton_thread_worker{{}};
     co_await w.start();
 
     std::vector<ss::future<size_t>> vec;