/* * Copyright (c) Meta Platforms, Inc. and affiliates. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include #include #include #include #include #include #include #include #include #include namespace apache { namespace thrift { namespace concurrency { class SFQThreadManagerConfig { public: SFQThreadManagerConfig() = default; SFQThreadManagerConfig& setPerturbInterval(std::chrono::milliseconds p) { perturb_ = p; return *this; } std::chrono::milliseconds getPerturbInterval() const { return perturb_; } // Currently only supporting fair queuing for upstream source. SFQThreadManagerConfig& setNumFairQueuesForUpstream(size_t numQueues) { numQueues_ = numQueues; return *this; } size_t getNumFairQueuesForUpstream() const { return numQueues_; } SFQThreadManagerConfig& setExecutors( std::array, N_PRIORITIES> executors) { executors_ = std::move(executors); return *this; } std::array, N_PRIORITIES>& getExecutors() { return executors_; } private: std::chrono::milliseconds perturb_{std::chrono::seconds(30)}; size_t numQueues_{1}; std::array, N_PRIORITIES> executors_; }; /** * Stochastic fair queue thread manager (SFQTM) * ============================================ * * This TM variant aims to address fairness concerns in multi-tenant systems by * fronting the CPU executor' UPSTREAM queue with a configurable number of * sub-queues used to isolate different request "tenants". Usage of the SFQTM * relies on users to identify different request tenants via encoding the data * in the priority provided to getKeepAlive() calls for UPSTREAM sources. * * Fairness in concurrency * ----------------------- * Fairness in the SFQTM context means that each tenant is allowed the same * number of tasks to reside on the internal executor queue. As an example, * let's say we have 2 tenants (A and B). When the internal executor's threads * are busy running tasks and begins queuing tasks, its queue will contain at * most a single task from A and a single task from B. As tasks from each tenant * are scheduled on a thread, a new task for that tenant is pulled from the * tenant's queue and placed in the internal executor's queue. * * Stochastic fairness * ------------------- * For practical reasons related to memory utilization and performance, multiple * tenants can potentially map to a single queue. When instantiating the * SFQTM, we allocate a fixed number of queues and hash the tenant ID onto this * fixed set of queues. This means that there is potential for multiple tenants * to map onto the same queue, resulting in a shared request concurrency * entitlement between them and unfairness. We mitigate this via perturbing the * hash function periodically via 'perturb'. This ensures that even if there is * a collision between two tenants, the condition will not be persistent. This * perturbation may result in request reordering for a tenant under heavy load. * * Notes * ~~~~~ * ** One may disable perturbing of the tenant hash by setting the period to * 0 seconds. * * ** If unspecified in the execution scope, tenant ID is assumed to be * zero. */ class SFQThreadManager : public ThreadManagerExecutorAdapter { public: explicit SFQThreadManager(SFQThreadManagerConfig config); ~SFQThreadManager() override; [[nodiscard]] KeepAlive<> getKeepAlive( ExecutionScope es, Source source) const override; [[noreturn]] void add( std::shared_ptr, int64_t, int64_t, ThreadManager::Source) noexcept override { LOG(FATAL) << "add*() is unsupported in SFQThreadManager, use getKeepAlive()"; } // Return size of tenantQueue for a given priority and tenantId size_t getTaskCount(ExecutionScope es); size_t pendingUpstreamTaskCount() const override; private: using ExecutorPtr = std::unique_ptr; void initPerturbation() { perturbationSchedule_.addFunction( [this]() mutable { perturbVal_.fetch_add(1, std::memory_order_relaxed); }, config_.getPerturbInterval(), "sfq_perturb"); perturbationSchedule_.start(); } uint64_t perturbId(uint64_t tenant, size_t val) const { return folly::hash::hash_combine(tenant, val); } // Returns a metered executor associated with a tenant ID. // // Since we use a stochastic fair queue, we have a fixed number of metered // queues and it's possible for multiple tenant IDs to collide on the same // metered queue which manifests in unfairness for those tenants. We // mitigate the issue by hashing the tenant IDs here with a periodically // changing value so that it is unlikely to persist beyond the current // period. folly::MeteredExecutor* getMeteredExecutor( size_t pri, uint64_t tenantId) const { const size_t p = perturbVal_.load(std::memory_order_relaxed); const uint64_t perturbedTenantId = perturbId(tenantId, p); CHECK_LT(pri, fqs_.size()); const size_t perturbedIdx = perturbedTenantId % fqs_[pri].size(); return fqs_[pri][perturbedIdx].get(); } // Set up the metered executors to act as fair queues. void initQueues(); SFQThreadManagerConfig config_; using MeteredExVec = std::vector>; std::array fqs_; std::atomic perturbVal_{0}; folly::FunctionScheduler perturbationSchedule_; }; } // namespace concurrency } // namespace thrift } // namespace apache