ydb/library/cpp/actors/core/balancer.cpp

#include "balancer.h"

#include "probes.h"

#include <library/cpp/actors/util/cpu_load_log.h>
#include <library/cpp/actors/util/datetime.h>
#include <library/cpp/actors/util/intrinsics.h>

#include <util/system/spinlock.h>

#include <algorithm>

namespace NActors {
    LWTRACE_USING(ACTORLIB_PROVIDER);

    // Describes balancing-related state of pool, the most notable is `Importance` to add new cpu
    struct TLevel {
        // Balancer will try to give more cpu to overloaded pools
        enum ELoadClass {
            Underloaded = 0,
            Moderate = 1,
            Overloaded = 2,
        };

        double ScaleFactor;
        ELoadClass LoadClass;
        ui64 Importance; // pool with lower importance is allowed to pass cpu to pool with higher, but the opposite is forbidden

        TLevel() {}

        TLevel(const TBalancingConfig& cfg, TPoolId poolId, ui64 currentCpus, double cpuIdle, ui64 addLatencyUs, ui64 worstLatencyUs) {
            ScaleFactor = double(currentCpus) / cfg.Cpus;
            if ((worstLatencyUs + addLatencyUs) < 2000 && cpuIdle > 1.0) { // Uderload criterion, based on estimated latency w/o 1 cpu
                LoadClass = Underloaded;
            } else if (worstLatencyUs > 2000 || cpuIdle < 0.2) { // Overload criterion, based on latency
                LoadClass = Overloaded;
            } else {
                LoadClass = Moderate;
            }
            Importance = MakeImportance(LoadClass, cfg.Priority, ScaleFactor, cpuIdle, poolId);
        }

    private:
        // Importance is simple ui64 value (from highest to lowest):
        //   2 Bits: LoadClass
        //   8 Bits: Priority
        //  10 Bits: -ScaleFactor (for max-min fairness with weights equal to TBalancingConfig::Cpus)
        //  10 Bits: -CpuIdle
        //   6 Bits: PoolId
        static ui64 MakeImportance(ELoadClass load, ui8 priority, double scaleFactor, double cpuIdle, TPoolId poolId) {
            ui64 idle = std::clamp<i64>(1024 - cpuIdle * 512, 0, 1023);
            ui64 scale = std::clamp<i64>(1024 - scaleFactor * 32, 0, 1023);

            Y_ABORT_UNLESS(ui64(load)     < (1ull << 2ull));
            Y_ABORT_UNLESS(ui64(priority) < (1ull << 8ull));
            Y_ABORT_UNLESS(ui64(scale)    < (1ull << 10ull));
            Y_ABORT_UNLESS(ui64(idle)     < (1ull << 10ull));
            Y_ABORT_UNLESS(ui64(poolId)   < (1ull << 6ull));

            static_assert(ui64(MaxPools) <= (1ull << 6ull));

            ui64 importance =
                (ui64(load)     << ui64(6 + 10 + 10 + 8)) |
                (ui64(priority) << ui64(6 + 10 + 10)) |
                (ui64(scale)    << ui64(6 + 10)) |
                (ui64(idle)     << ui64(6)) |
                ui64(poolId);
            return importance;
        }
    };

    // Main balancer implemenation
    class TBalancer: public IBalancer {
    private:
        struct TCpu;
        struct TPool;

        bool Disabled = true;
        TSpinLock Lock;
        ui64 NextBalanceTs;
        TVector<TCpu> Cpus; // Indexed by CpuId, can have gaps
        TVector<TPool> Pools; // Indexed by PoolId, can have gaps
        TBalancerConfig Config;

    public:

        ui64 GetPeriodUs() override;
        // Setup
        TBalancer(const TBalancerConfig& config, const TVector<TUnitedExecutorPoolConfig>& unitedPools, ui64 ts);
        bool AddCpu(const TCpuAllocation& cpuAlloc, TCpuState* cpu) override;
        ~TBalancer();

        // Balancing
        bool TryLock(ui64 ts) override;
        void SetPoolStats(TPoolId pool, const TBalancerStats& stats) override;
        void Balance() override;
        void Unlock() override;

    private:
        void MoveCpu(TPool& from, TPool& to);
    };

    struct TBalancer::TPool {
        TBalancingConfig Config;
        TPoolId PoolId;
        TString PoolName;

        // Input data for balancing
        TBalancerStats Prev;
        TBalancerStats Next;

        // Derived stats
        double CpuLoad;
        double CpuIdle;

        // Classification
        // NOTE: We want to avoid passing cpu back and forth, so we must consider not only current level,
        // NOTE: but expected levels after movements also
        TLevel CurLevel; // Level with current amount of cpu
        TLevel AddLevel; // Level after one cpu acception
        TLevel SubLevel; // Level after one cpu donation

        // Balancing state
        ui64 CurrentCpus = 0; // Total number of cpus assigned for this pool (zero means pools is not balanced)
        ui64 PrevCpus = 0; // Cpus in last period

        explicit TPool(const TBalancingConfig& cfg = {})
            : Config(cfg)
        {}

        void Configure(const TBalancingConfig& cfg, const TString& poolName) {
            Config = cfg;
            // Enforce constraints
            if (Config.Cpus > 0) {
                Config.MinCpus = std::clamp<ui32>(Config.MinCpus, 1, Config.Cpus);
                Config.MaxCpus = Max<ui32>(Config.MaxCpus, Config.Cpus);
            } else {
                Y_ABORT_UNLESS(Config.Cpus == 0,
                        "Unexpected negative Config.Cpus# %" PRIi64,
                        (i64)Config.Cpus);
                Config.MinCpus = 0;
                Config.MaxCpus = 0;
            }
            PoolName = poolName;
        }
    };

    struct TBalancer::TCpu {
        TCpuState* State = nullptr; // Cpu state, nullptr means cpu is not used (gap)
        TCpuAllocation Alloc;
        TPoolId Current;
        TPoolId Assigned;
    };

    TBalancer::TBalancer(const TBalancerConfig& config, const TVector<TUnitedExecutorPoolConfig>& unitedPools, ui64 ts)
        : NextBalanceTs(ts)
        , Config(config)
    {
        for (TPoolId pool = 0; pool < MaxPools; pool++) {
            Pools.emplace_back();
            Pools.back().PoolId = pool;
        }
        for (const TUnitedExecutorPoolConfig& united : unitedPools) {
            Pools[united.PoolId].Configure(united.Balancing, united.PoolName);
        }
    }

    TBalancer::~TBalancer() {
    }

    bool TBalancer::AddCpu(const TCpuAllocation& cpuAlloc, TCpuState* state) {
        // Setup
        TCpuId cpuId = cpuAlloc.CpuId;
        if (Cpus.size() <= cpuId) {
            Cpus.resize(cpuId + 1);
        }
        TCpu& cpu = Cpus[cpuId];
        cpu.State = state;
        cpu.Alloc = cpuAlloc;

        // Fill every pool with cpus up to TBalancingConfig::Cpus
        TPoolId pool = 0;
        for (TPool& p : Pools) {
            if (p.CurrentCpus < p.Config.Cpus) {
                p.CurrentCpus++;
                break;
            }
            pool++;
        }
        if (pool != MaxPools) { // cpu under balancer control
            state->SwitchPool(pool);
            state->AssignPool(pool);
            Disabled = false;
            return true;
        }
        return false; // non-balanced cpu
    }

    bool TBalancer::TryLock(ui64 ts) {
        if (!Disabled && NextBalanceTs < ts && Lock.TryAcquire()) {
            NextBalanceTs = ts + Us2Ts(Config.PeriodUs);
            return true;
        }
        return false;
    }

    void TBalancer::SetPoolStats(TPoolId pool, const TBalancerStats& stats) {
        Y_ABORT_UNLESS(pool < MaxPools);
        TPool& p = Pools[pool];
        p.Prev = p.Next;
        p.Next = stats;
    }

    void TBalancer::Balance() {
        // Update every cpu state
        for (TCpu& cpu : Cpus) {
            if (cpu.State) {
                cpu.State->Load(cpu.Assigned, cpu.Current);
                if (cpu.Current < MaxPools && cpu.Current != cpu.Assigned) {
                    return; // previous movement has not been applied yet, wait
                }
            }
        }

        // Process stats, classify and compute pool importance
        TStackVec<TPool*, MaxPools> order;
        for (TPool& pool : Pools) {
            if (pool.Config.Cpus == 0) {
                continue; // skip gaps (non-existent or non-united pools)
            }
            if (pool.Prev.Ts == 0 || pool.Prev.Ts >= pool.Next.Ts) {
                return; // invalid stats
            }

            // Compute derived stats
            pool.CpuLoad = (pool.Next.CpuUs - pool.Prev.CpuUs) / Ts2Us(pool.Next.Ts - pool.Prev.Ts);
            if (pool.Prev.IdleUs == ui64(-1) || pool.Next.IdleUs == ui64(-1)) {
                pool.CpuIdle = pool.CurrentCpus - pool.CpuLoad; // for tests
            } else {
                pool.CpuIdle = (pool.Next.IdleUs - pool.Prev.IdleUs) / Ts2Us(pool.Next.Ts - pool.Prev.Ts);
            }

            // Compute levels
            pool.CurLevel = TLevel(pool.Config, pool.PoolId, pool.CurrentCpus, pool.CpuIdle,
                    pool.Next.ExpectedLatencyIncreaseUs, pool.Next.WorstActivationTimeUs);
            pool.AddLevel = TLevel(pool.Config, pool.PoolId, pool.CurrentCpus + 1, pool.CpuIdle,
                    0, pool.Next.WorstActivationTimeUs); // we expect taken cpu to became utilized
            pool.SubLevel = TLevel(pool.Config, pool.PoolId, pool.CurrentCpus - 1, pool.CpuIdle - 1,
                    pool.Next.ExpectedLatencyIncreaseUs, pool.Next.WorstActivationTimeUs);

            // Prepare for balancing
            pool.PrevCpus = pool.CurrentCpus;
            order.push_back(&pool);
        }

        // Sort pools by importance
        std::sort(order.begin(), order.end(), [] (TPool* l, TPool* r) {return l->CurLevel.Importance < r->CurLevel.Importance; });
        for (TPool* pool : order) {
            LWPROBE(PoolStats, pool->PoolId, pool->PoolName, pool->CurrentCpus, pool->CurLevel.LoadClass, pool->Config.Priority, pool->CurLevel.ScaleFactor, pool->CpuIdle, pool->CpuLoad, pool->CurLevel.Importance, pool->AddLevel.Importance, pool->SubLevel.Importance);
        }

        // Move cpus from lower importance to higher importance pools
        for (auto toIter = order.rbegin(); toIter != order.rend(); ++toIter) {
            TPool& to = **toIter;
            if (to.CurLevel.LoadClass == TLevel::Overloaded && // if pool is overloaded
                to.CurrentCpus < to.Config.MaxCpus) // and constraints would not be violated
            {
                for (auto fromIter = order.begin(); (*fromIter)->CurLevel.Importance < to.CurLevel.Importance; ++fromIter) {
                    TPool& from = **fromIter;
                    if (from.CurrentCpus == from.PrevCpus && // if not balanced yet
                        from.CurrentCpus > from.Config.MinCpus && // and constraints would not be violated
                        from.SubLevel.Importance <= to.AddLevel.Importance) // and which of two pools is more important would not change after cpu movement
                    {
                        MoveCpu(from, to);
                        from.CurrentCpus--;
                        to.CurrentCpus++;
                        break;
                    }
                }
            }
        }
    }

    void TBalancer::MoveCpu(TBalancer::TPool& from, TBalancer::TPool& to) {
        for (auto ci = Cpus.rbegin(), ce = Cpus.rend(); ci != ce; ci++) {
            TCpu& cpu = *ci;
            if (!cpu.State) {
                continue;
            }
            if (cpu.Assigned == from.PoolId) {
                cpu.State->AssignPool(to.PoolId);
                cpu.Assigned = to.PoolId;
                LWPROBE(MoveCpu, from.PoolId, to.PoolId, from.PoolName, to.PoolName, cpu.Alloc.CpuId);
                return;
            }
        }
        Y_ABORT();
    }

    void TBalancer::Unlock() {
        Lock.Release();
    }

    ui64 TBalancer::GetPeriodUs() {
        return Config.PeriodUs;
    }

    IBalancer* MakeBalancer(const TBalancerConfig& config, const TVector<TUnitedExecutorPoolConfig>& unitedPools, ui64 ts) {
        return new TBalancer(config, unitedPools, ts);
    }
}