tursom/memgraph
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #657 from memgraph/tyler_simulator_determinism_test

Browse Source 
Improve simulator determinism
This commit is contained in:
János Benjamin Antal 2022-11-23 09:14:25 +01:00 committed by GitHub
commit d926c4c4a4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
20 changed files with 393 additions and 129 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
src/coordinator/coordinator_worker.hpp
View File

@ -71,6 +71,9 @@ struct QueueInner {
// starvation by sometimes randomizing priorities, rather than following a strict
// prioritization.
std::deque<Message> queue;
uint64_t submitted = 0;
uint64_t calls_to_pop = 0;
};
/// There are two reasons to implement our own Queue instead of using
@ -86,6 +89,8 @@ class Queue {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
inner_->submitted++;
inner_->queue.emplace_back(std::move(message));
} // lock dropped before notifying condition variable
@ -96,6 +101,9 @@ class Queue {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
inner_->calls_to_pop++;
inner_->cv.notify_all();
while (inner_->queue.empty()) {
inner_->cv.wait(lock);
}
@ -105,6 +113,15 @@ class Queue {
return message;
}
void BlockOnQuiescence() const {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
while (inner_->calls_to_pop <= inner_->submitted) {
inner_->cv.wait(lock);
}
}
};
/// A CoordinatorWorker owns Raft<CoordinatorRsm> instances. receives messages from the MachineManager.
@ -129,9 +146,7 @@ class CoordinatorWorker {
public:
CoordinatorWorker(io::Io<IoImpl> io, Queue queue, Coordinator coordinator)
: io_(std::move(io)),
queue_(std::move(queue)),
coordinator_{std::move(io_.ForkLocal()), {}, std::move(coordinator)} {}
: io_(std::move(io)), queue_(std::move(queue)), coordinator_{std::move(io_), {}, std::move(coordinator)} {}
CoordinatorWorker(CoordinatorWorker &&) noexcept = default;
CoordinatorWorker &operator=(CoordinatorWorker &&) noexcept = default;
@ -140,15 +155,12 @@ class CoordinatorWorker {
~CoordinatorWorker() = default;
void Run() {
while (true) {
bool should_continue = true;
while (should_continue) {
Message message = queue_.Pop();
const bool should_continue = std::visit(
[this](auto &&msg) { return this->Process(std::forward<decltype(msg)>(msg)); }, std::move(message));
if (!should_continue) {
return;
}
should_continue = std::visit([this](auto &&msg) { return this->Process(std::forward<decltype(msg)>(msg)); },
std::move(message));
}
}
};

28
src/coordinator/shard_map.cpp
View File

@ -228,7 +228,7 @@ Hlc ShardMap::IncrementShardMapVersion() noexcept {
return shard_map_version;
}
// TODO(antaljanosbenjamin) use a single map for all name id
// TODO(antaljanosbenjamin) use a single map for all name id
// mapping and a single counter to maintain the next id
std::unordered_map<uint64_t, std::string> ShardMap::IdToNames() {
std::unordered_map<uint64_t, std::string> id_to_names;
@ -248,6 +248,25 @@ std::unordered_map<uint64_t, std::string> ShardMap::IdToNames() {
Hlc ShardMap::GetHlc() const noexcept { return shard_map_version; }
boost::uuids::uuid NewShardUuid(uint64_t shard_id) {
return boost::uuids::uuid{0,
0,
0,
0,
0,
0,
0,
0,
static_cast<unsigned char>(shard_id >> 56U),
static_cast<unsigned char>(shard_id >> 48U),
static_cast<unsigned char>(shard_id >> 40U),
static_cast<unsigned char>(shard_id >> 32U),
static_cast<unsigned char>(shard_id >> 24U),
static_cast<unsigned char>(shard_id >> 16U),
static_cast<unsigned char>(shard_id >> 8U),
static_cast<unsigned char>(shard_id)};
}
std::vector<ShardToInitialize> ShardMap::AssignShards(Address storage_manager,
std::set<boost::uuids::uuid> initialized) {
std::vector<ShardToInitialize> ret{};
@ -268,6 +287,7 @@ std::vector<ShardToInitialize> ShardMap::AssignShards(Address storage_manager,
if (initialized.contains(aas.address.unique_id)) {
machine_contains_shard = true;
if (aas.status != Status::CONSENSUS_PARTICIPANT) {
mutated = true;
spdlog::info("marking shard as full consensus participant: {}", aas.address.unique_id);
aas.status = Status::CONSENSUS_PARTICIPANT;
}
@ -292,10 +312,13 @@ std::vector<ShardToInitialize> ShardMap::AssignShards(Address storage_manager,
}
if (!machine_contains_shard && shard.size() < label_space.replication_factor) {
// increment version for each new uuid for deterministic creation
IncrementShardMapVersion();
Address address = storage_manager;
// TODO(tyler) use deterministic UUID so that coordinators don't diverge here
address.unique_id = boost::uuids::uuid{boost::uuids::random_generator()()},
address.unique_id = NewShardUuid(shard_map_version.logical_id);
spdlog::info("assigning shard manager to shard");
@ -383,6 +406,7 @@ std::optional<LabelId> ShardMap::InitializeNewLabel(std::string label_name, std:
void ShardMap::AddServer(Address server_address) {
// Find a random place for the server to plug in
}
std::optional<LabelId> ShardMap::GetLabelId(const std::string &label) const {
if (const auto it = labels.find(label); it != labels.end()) {
return it->second;

25
src/io/address.hpp
View File

@ -20,6 +20,8 @@
#include <boost/uuid/uuid_generators.hpp>
#include <boost/uuid/uuid_io.hpp>
#include "utils/logging.hpp"
namespace memgraph::io {
struct PartialAddress {
@ -58,18 +60,39 @@ struct Address {
uint16_t last_known_port;
static Address TestAddress(uint16_t port) {
MG_ASSERT(port <= 255);
return Address{
.unique_id = boost::uuids::uuid{boost::uuids::random_generator()()},
.unique_id = boost::uuids::uuid{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, static_cast<unsigned char>(port)},
.last_known_port = port,
};
}
// NB: don't use this in test code because it is non-deterministic
static Address UniqueLocalAddress() {
return Address{
.unique_id = boost::uuids::uuid{boost::uuids::random_generator()()},
};
}
/// `Coordinator`s have constant UUIDs because there is at most one per ip/port pair.
Address ForkLocalCoordinator() {
return Address{
.unique_id = boost::uuids::uuid{1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
.last_known_ip = last_known_ip,
.last_known_port = last_known_port,
};
}
/// `ShardManager`s have constant UUIDs because there is at most one per ip/port pair.
Address ForkLocalShardManager() {
return Address{
.unique_id = boost::uuids::uuid{2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
.last_known_ip = last_known_ip,
.last_known_port = last_known_port,
};
}
/// Returns a new ID with the same IP and port but a unique UUID.
Address ForkUniqueAddress() {
return Address{

6
src/io/future.hpp
View File

@ -64,7 +64,6 @@ class Shared {
waiting_ = true;
while (!item_) {
bool simulator_progressed = false;
if (simulator_notifier_) [[unlikely]] {
// We can't hold our own lock while notifying
// the simulator because notifying the simulator
@ -77,7 +76,7 @@ class Shared {
// so we have to get out of its way to avoid
// a cyclical deadlock.
lock.unlock();
simulator_progressed = std::invoke(simulator_notifier_);
std::invoke(simulator_notifier_);
lock.lock();
if (item_) {
// item may have been filled while we
@ -85,8 +84,7 @@ class Shared {
// the simulator of our waiting_ status.
break;
}
}
if (!simulator_progressed) [[likely]] {
} else {
cv_.wait(lock);
}
MG_ASSERT(!consumed_, "Future consumed twice!");

5
src/io/message_histogram_collector.hpp
View File

@ -13,6 +13,7 @@
#include <chrono>
#include <cmath>
#include <compare>
#include <unordered_map>
#include <boost/core/demangle.hpp>
@ -39,6 +40,8 @@ struct LatencyHistogramSummary {
Duration p100;
Duration sum;
friend bool operator==(const LatencyHistogramSummary &lhs, const LatencyHistogramSummary &rhs) = default;
friend std::ostream &operator<<(std::ostream &in, const LatencyHistogramSummary &histo) {
in << "{ \"count\": " << histo.count;
in << ", \"p0\": " << histo.p0.count();
@ -80,6 +83,8 @@ struct LatencyHistogramSummaries {
return output;
}
friend bool operator==(const LatencyHistogramSummaries &lhs, const LatencyHistogramSummaries &rhs) = default;
friend std::ostream &operator<<(std::ostream &in, const LatencyHistogramSummaries &histo) {
using memgraph::utils::print_helpers::operator<<;
in << histo.latencies;

5
src/io/rsm/raft.hpp
View File

@ -19,7 +19,6 @@
#include <map>
#include <set>
#include <thread>
#include <unordered_map>
#include <vector>
#include <boost/core/demangle.hpp>
@ -182,7 +181,7 @@ struct PendingClientRequest {
struct Leader {
std::map<Address, FollowerTracker> followers;
std::unordered_map<LogIndex, PendingClientRequest> pending_client_requests;
std::map<LogIndex, PendingClientRequest> pending_client_requests;
Time last_broadcast = Time::min();
std::string static ToString() { return "\tLeader \t"; }
@ -683,7 +682,7 @@ class Raft {
return Leader{
.followers = std::move(followers),
.pending_client_requests = std::unordered_map<LogIndex, PendingClientRequest>(),
.pending_client_requests = std::map<LogIndex, PendingClientRequest>(),
};
}

83
src/io/simulator/simulator_handle.cpp
View File

@ -23,6 +23,12 @@ namespace memgraph::io::simulator {
void SimulatorHandle::ShutDown() {
std::unique_lock<std::mutex> lock(mu_);
should_shut_down_ = true;
for (auto it = promises_.begin(); it != promises_.end();) {
auto &[promise_key, dop] = *it;
std::move(dop).promise.TimeOut();
it = promises_.erase(it);
}
can_receive_.clear();
cv_.notify_all();
}
@ -46,52 +52,84 @@ void SimulatorHandle::IncrementServerCountAndWaitForQuiescentState(Address addre
const bool all_servers_blocked = blocked_servers == server_addresses_.size();
if (all_servers_blocked) {
spdlog::trace("quiescent state detected - {} out of {} servers now blocked on receive", blocked_servers,
server_addresses_.size());
return;
}
spdlog::trace("not returning from quiescent because we see {} blocked out of {}", blocked_servers,
server_addresses_.size());
cv_.wait(lock);
}
}
bool SortInFlight(const std::pair<Address, OpaqueMessage> &lhs, const std::pair<Address, OpaqueMessage> &rhs) {
// NB: never sort based on the request ID etc...
// This should only be used from std::stable_sort
// because by comparing on the from_address alone,
// we expect the sender ordering to remain
// deterministic.
const auto &[addr_1, om_1] = lhs;
const auto &[addr_2, om_2] = rhs;
return om_1.from_address < om_2.from_address;
}
bool SimulatorHandle::MaybeTickSimulator() {
std::unique_lock<std::mutex> lock(mu_);
const size_t blocked_servers = blocked_on_receive_.size();
if (blocked_servers < server_addresses_.size()) {
if (should_shut_down_ || blocked_servers < server_addresses_.size()) {
// we only need to advance the simulator when all
// servers have reached a quiescent state, blocked
// on their own futures or receive methods.
return false;
}
// We allow the simulator to progress the state of the system only
// after all servers are blocked on receive.
spdlog::trace("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ simulator tick ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~");
stats_.simulator_ticks++;
blocked_on_receive_.clear();
cv_.notify_all();
TimeoutPromisesPastDeadline();
bool timed_anything_out = TimeoutPromisesPastDeadline();
if (timed_anything_out) {
spdlog::trace("simulator progressing: timed out a request");
}
const Duration clock_advance = std::chrono::microseconds{time_distrib_(rng_)};
// We don't always want to advance the clock with every message that we deliver because
// when we advance it for every message, it causes timeouts to occur for any "request path"
// over a certain length. Alternatively, we don't want to simply deliver multiple messages
// in a single simulator tick because that would reduce the amount of concurrent message
// mixing that may naturally occur in production. This approach is to mod the random clock
// advance by a prime number (hopefully avoiding most harmonic effects that would be introduced
// by only advancing the clock by an even amount etc...) and only advancing the clock close to
// half of the time.
if (clock_advance.count() % 97 > 49) {
spdlog::trace("simulator progressing: clock advanced by {}", clock_advance.count());
cluster_wide_time_microseconds_ += clock_advance;
stats_.elapsed_time = cluster_wide_time_microseconds_ - config_.start_time;
}
if (cluster_wide_time_microseconds_ >= config_.abort_time) {
spdlog::error(
"Cluster has executed beyond its configured abort_time, and something may be failing to make progress "
"in an expected amount of time. The SimulatorConfig.rng_seed for this run is {}",
config_.rng_seed);
throw utils::BasicException{"Cluster has executed beyond its configured abort_time"};
}
if (in_flight_.empty()) {
// return early here because there are no messages to schedule
// We tick the clock forward when all servers are blocked but
// there are no in-flight messages to schedule delivery of.
const Duration clock_advance = std::chrono::microseconds{time_distrib_(rng_)};
cluster_wide_time_microseconds_ += clock_advance;
if (cluster_wide_time_microseconds_ >= config_.abort_time) {
if (should_shut_down_) {
return false;
}
spdlog::error(
"Cluster has executed beyond its configured abort_time, and something may be failing to make progress "
"in an expected amount of time.");
throw utils::BasicException{"Cluster has executed beyond its configured abort_time"};
}
return true;
}
if (config_.scramble_messages) {
std::stable_sort(in_flight_.begin(), in_flight_.end(), SortInFlight);
if (config_.scramble_messages && in_flight_.size() > 1) {
// scramble messages
std::uniform_int_distribution<size_t> swap_distrib(0, in_flight_.size() - 1);
const size_t swap_index = swap_distrib(rng_);
@ -120,17 +158,22 @@ bool SimulatorHandle::MaybeTickSimulator() {
if (should_drop || normal_timeout) {
stats_.timed_out_requests++;
dop.promise.TimeOut();
spdlog::trace("simulator timing out request ");
} else {
stats_.total_responses++;
Duration response_latency = cluster_wide_time_microseconds_ - dop.requested_at;
auto type_info = opaque_message.type_info;
dop.promise.Fill(std::move(opaque_message), response_latency);
histograms_.Measure(type_info, response_latency);
spdlog::trace("simulator replying to request");
}
} else if (should_drop) {
// don't add it anywhere, let it drop
spdlog::trace("simulator silently dropping request");
} else {
// add to can_receive_ if not
spdlog::trace("simulator adding message to can_receive_ from {} to {}", opaque_message.from_address.last_known_port,
opaque_message.to_address.last_known_port);
const auto &[om_vec, inserted] =
can_receive_.try_emplace(to_address.ToPartialAddress(), std::vector<OpaqueMessage>());
om_vec->second.emplace_back(std::move(opaque_message));

48
src/io/simulator/simulator_handle.hpp
View File

@ -52,26 +52,29 @@ class SimulatorHandle {
std::set<Address> blocked_on_receive_;
std::set<Address> server_addresses_;
std::mt19937 rng_;
std::uniform_int_distribution<int> time_distrib_{5, 50};
std::uniform_int_distribution<int> time_distrib_{0, 30000};
std::uniform_int_distribution<int> drop_distrib_{0, 99};
SimulatorConfig config_;
MessageHistogramCollector histograms_;
RequestId request_id_counter_{0};
void TimeoutPromisesPastDeadline() {
bool TimeoutPromisesPastDeadline() {
bool timed_anything_out = false;
const Time now = cluster_wide_time_microseconds_;
for (auto it = promises_.begin(); it != promises_.end();) {
auto &[promise_key, dop] = *it;
if (dop.deadline < now && config_.perform_timeouts) {
spdlog::info("timing out request from requester {}.", promise_key.requester_address.ToString());
spdlog::trace("timing out request from requester {}.", promise_key.requester_address.ToString());
std::move(dop).promise.TimeOut();
it = promises_.erase(it);
stats_.timed_out_requests++;
timed_anything_out = true;
} else {
++it;
}
}
return timed_anything_out;
}
public:
@ -103,6 +106,7 @@ class SimulatorHandle {
template <Message Request, Message Response>
ResponseFuture<Response> SubmitRequest(Address to_address, Address from_address, Request &&request, Duration timeout,
std::function<bool()> &&maybe_tick_simulator) {
spdlog::trace("submitting request to {}", to_address.last_known_port);
auto type_info = TypeInfoFor(request);
auto [future, promise] = memgraph::io::FuturePromisePairWithNotifier<ResponseResult<Response>>(
@ -146,8 +150,6 @@ class SimulatorHandle {
requires(sizeof...(Ms) > 0) RequestResult<Ms...> Receive(const Address &receiver, Duration timeout) {
std::unique_lock<std::mutex> lock(mu_);
blocked_on_receive_.emplace(receiver);
const Time deadline = cluster_wide_time_microseconds_ + timeout;
auto partial_address = receiver.ToPartialAddress();
@ -164,38 +166,40 @@ class SimulatorHandle {
auto m_opt = std::move(message).Take<Ms...>();
MG_ASSERT(m_opt.has_value(), "Wrong message type received compared to the expected type");
blocked_on_receive_.erase(receiver);
return std::move(m_opt).value();
}
}
lock.unlock();
bool made_progress = MaybeTickSimulator();
lock.lock();
if (!should_shut_down_ && !made_progress) {
if (!should_shut_down_) {
if (!blocked_on_receive_.contains(receiver)) {
blocked_on_receive_.emplace(receiver);
spdlog::trace("blocking receiver {}", receiver.ToPartialAddress().port);
cv_.notify_all();
}
cv_.wait(lock);
}
}
blocked_on_receive_.erase(receiver);
spdlog::trace("timing out receiver {}", receiver.ToPartialAddress().port);
return TimedOut{};
}
template <Message M>
void Send(Address to_address, Address from_address, RequestId request_id, M message) {
spdlog::trace("sending message from {} to {}", from_address.last_known_port, to_address.last_known_port);
auto type_info = TypeInfoFor(message);
std::unique_lock<std::mutex> lock(mu_);
std::any message_any(std::move(message));
OpaqueMessage om{.to_address = to_address,
.from_address = from_address,
.request_id = request_id,
.message = std::move(message_any),
.type_info = type_info};
in_flight_.emplace_back(std::make_pair(std::move(to_address), std::move(om)));
{
std::unique_lock<std::mutex> lock(mu_);
std::any message_any(std::move(message));
OpaqueMessage om{.to_address = to_address,
.from_address = from_address,
.request_id = request_id,
.message = std::move(message_any),
.type_info = type_info};
in_flight_.emplace_back(std::make_pair(std::move(to_address), std::move(om)));
stats_.total_messages++;
stats_.total_messages++;
} // lock dropped before cv notification
cv_.notify_all();
}

21
src/io/simulator/simulator_stats.hpp
View File

@ -13,6 +13,10 @@
#include <cstdint>
#include <fmt/format.h>
#include "io/time.hpp"
namespace memgraph::io::simulator {
struct SimulatorStats {
uint64_t total_messages = 0;
@ -21,5 +25,22 @@ struct SimulatorStats {
uint64_t total_requests = 0;
uint64_t total_responses = 0;
uint64_t simulator_ticks = 0;
Duration elapsed_time;
friend bool operator==(const SimulatorStats & /* lhs */, const SimulatorStats & /* rhs */) = default;
friend std::ostream &operator<<(std::ostream &in, const SimulatorStats &stats) {
auto elapsed_ms = std::chrono::duration_cast<std::chrono::milliseconds>(stats.elapsed_time).count();
std::string formated = fmt::format(
"SimulatorStats {{ total_messages: {}, dropped_messages: {}, timed_out_requests: {}, total_requests: {}, "
"total_responses: {}, simulator_ticks: {}, elapsed_time: {}ms }}",
stats.total_messages, stats.dropped_messages, stats.timed_out_requests, stats.total_requests,
stats.total_responses, stats.simulator_ticks, elapsed_ms);
in << formated;
return in;
}
};
}; // namespace memgraph::io::simulator

9
src/io/transport.hpp
View File

@ -137,7 +137,14 @@ class Io {
Address GetAddress() { return address_; }
void SetAddress(Address address) { address_ = address; }
Io<I> ForkLocal() { return Io(implementation_, address_.ForkUniqueAddress()); }
Io<I> ForkLocal(boost::uuids::uuid uuid) {
Address new_address{
.unique_id = uuid,
.last_known_ip = address_.last_known_ip,
.last_known_port = address_.last_known_port,
};
return Io(implementation_, new_address);
}
LatencyHistogramSummaries ResponseLatencies() { return implementation_.ResponseLatencies(); }
};

1
src/machine_manager/machine_config.hpp
View File

@ -42,6 +42,7 @@ struct MachineConfig {
boost::asio::ip::address listen_ip;
uint16_t listen_port;
size_t shard_worker_threads = std::max(static_cast<unsigned int>(1), std::thread::hardware_concurrency());
bool sync_message_handling = false;
};
} // namespace memgraph::machine_manager

41
src/machine_manager/machine_manager.hpp
View File

@ -78,10 +78,10 @@ class MachineManager {
MachineManager(io::Io<IoImpl> io, MachineConfig config, Coordinator coordinator)
: io_(io),
config_(config),
coordinator_address_(io.GetAddress().ForkUniqueAddress()),
shard_manager_{io.ForkLocal(), config.shard_worker_threads, coordinator_address_} {
auto coordinator_io = io.ForkLocal();
coordinator_io.SetAddress(coordinator_address_);
coordinator_address_(io.GetAddress().ForkLocalCoordinator()),
shard_manager_{io.ForkLocal(io.GetAddress().ForkLocalShardManager().unique_id), config.shard_worker_threads,
coordinator_address_} {
auto coordinator_io = io.ForkLocal(coordinator_address_.unique_id);
CoordinatorWorker coordinator_worker{coordinator_io, coordinator_queue_, coordinator};
coordinator_handle_ = std::jthread([coordinator = std::move(coordinator_worker)]() mutable { coordinator.Run(); });
}
@ -101,11 +101,23 @@ class MachineManager {
Address CoordinatorAddress() { return coordinator_address_; }
void Run() {
while (!io_.ShouldShutDown()) {
while (true) {
MaybeBlockOnSyncHandling();
if (io_.ShouldShutDown()) {
break;
}
const auto now = io_.Now();
uint64_t now_us = now.time_since_epoch().count();
uint64_t next_us = next_cron_.time_since_epoch().count();
if (now >= next_cron_) {
spdlog::info("now {} >= next_cron_ {}", now_us, next_us);
next_cron_ = Cron();
} else {
spdlog::info("now {} < next_cron_ {}", now_us, next_us);
}
Duration receive_timeout = std::max(next_cron_, now) - now;
@ -194,10 +206,27 @@ class MachineManager {
}
private:
// This method exists for controlling concurrency
// during deterministic simulation testing.
void MaybeBlockOnSyncHandling() {
if (!config_.sync_message_handling) {
return;
}
// block on coordinator
coordinator_queue_.BlockOnQuiescence();
// block on shards
shard_manager_.BlockOnQuiescence();
}
Time Cron() {
spdlog::info("running MachineManager::Cron, address {}", io_.GetAddress().ToString());
coordinator_queue_.Push(coordinator::coordinator_worker::Cron{});
return shard_manager_.Cron();
MaybeBlockOnSyncHandling();
Time ret = shard_manager_.Cron();
MaybeBlockOnSyncHandling();
return ret;
}
};

6
src/query/v2/interpreter.cpp
View File

@ -800,7 +800,11 @@ InterpreterContext::InterpreterContext(storage::v3::Shard *db, const Interpreter
Interpreter::Interpreter(InterpreterContext *interpreter_context) : interpreter_context_(interpreter_context) {
MG_ASSERT(interpreter_context_, "Interpreter context must not be NULL");
auto query_io = interpreter_context_->io.ForkLocal();
// TODO(tyler) make this deterministic so that it can be tested.
auto random_uuid = boost::uuids::uuid{boost::uuids::random_generator()()};
auto query_io = interpreter_context_->io.ForkLocal(random_uuid);
shard_request_manager_ = std::make_unique<msgs::ShardRequestManager<io::local_transport::LocalTransport>>(
coordinator::CoordinatorClient<io::local_transport::LocalTransport>(
query_io, interpreter_context_->coordinator_address, std::vector{interpreter_context_->coordinator_address}),

6
src/storage/v3/shard_manager.hpp
View File

@ -190,6 +190,12 @@ class ShardManager {
});
}
void BlockOnQuiescence() {
for (const auto &worker : workers_) {
worker.BlockOnQuiescence();
}
}
private:
io::Io<IoImpl> io_;
std::vector<shard_worker::Queue> workers_;

32
src/storage/v3/shard_worker.hpp
View File

@ -80,6 +80,9 @@ struct QueueInner {
// starvation by sometimes randomizing priorities, rather than following a strict
// prioritization.
std::deque<Message> queue;
uint64_t submitted = 0;
uint64_t calls_to_pop = 0;
};
/// There are two reasons to implement our own Queue instead of using
@ -95,6 +98,8 @@ class Queue {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
inner_->submitted++;
inner_->queue.emplace_back(std::forward<Message>(message));
} // lock dropped before notifying condition variable
@ -105,6 +110,9 @@ class Queue {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
inner_->calls_to_pop++;
inner_->cv.notify_all();
while (inner_->queue.empty()) {
inner_->cv.wait(lock);
}
@ -114,6 +122,15 @@ class Queue {
return message;
}
void BlockOnQuiescence() const {
MG_ASSERT(inner_.use_count() > 0);
std::unique_lock<std::mutex> lock(inner_->mu);
while (inner_->calls_to_pop <= inner_->submitted) {
inner_->cv.wait(lock);
}
}
};
/// A ShardWorker owns Raft<ShardRsm> instances. receives messages from the ShardManager.
@ -122,7 +139,6 @@ class ShardWorker {
io::Io<IoImpl> io_;
Queue queue_;
std::priority_queue<std::pair<Time, uuid>, std::vector<std::pair<Time, uuid>>, std::greater<>> cron_schedule_;
Time next_cron_ = Time::min();
std::map<uuid, ShardRaft<IoImpl>> rsm_map_;
bool Process(ShutDown && /* shut_down */) { return false; }
@ -175,10 +191,7 @@ class ShardWorker {
return;
}
auto rsm_io = io_.ForkLocal();
auto io_addr = rsm_io.GetAddress();
io_addr.unique_id = to_init.uuid;
rsm_io.SetAddress(io_addr);
auto rsm_io = io_.ForkLocal(to_init.uuid);
// TODO(tyler) get peers from Coordinator in HeartbeatResponse
std::vector<Address> rsm_peers = {};
@ -208,15 +221,12 @@ class ShardWorker {
~ShardWorker() = default;
void Run() {
while (true) {
bool should_continue = true;
while (should_continue) {
Message message = queue_.Pop();
const bool should_continue =
should_continue =
std::visit([&](auto &&msg) { return Process(std::forward<decltype(msg)>(msg)); }, std::move(message));
if (!should_continue) {
return;
}
}
}
};

55
tests/simulation/basic_request.cpp
View File

@ -9,17 +9,25 @@
// by the Apache License, Version 2.0, included in the file
// licenses/APL.txt.
#include <memory>
#include <thread>
#include <gmock/gmock.h>
#include <gtest/gtest.h>
#include <spdlog/cfg/env.h>
#include "io/message_histogram_collector.hpp"
#include "io/simulator/simulator.hpp"
#include "utils/print_helpers.hpp"
using memgraph::io::Address;
using memgraph::io::Io;
using memgraph::io::LatencyHistogramSummaries;
using memgraph::io::ResponseFuture;
using memgraph::io::ResponseResult;
using memgraph::io::simulator::Simulator;
using memgraph::io::simulator::SimulatorConfig;
using memgraph::io::simulator::SimulatorStats;
using memgraph::io::simulator::SimulatorTransport;
struct CounterRequest {
@ -40,6 +48,7 @@ void run_server(Io<SimulatorTransport> io) {
std::cout << "[SERVER] Error, continue" << std::endl;
continue;
}
std::cout << "[SERVER] Got message" << std::endl;
auto request_envelope = request_result.GetValue();
auto req = std::get<CounterRequest>(request_envelope.message);
@ -50,13 +59,7 @@ void run_server(Io<SimulatorTransport> io) {
}
}
int main() {
auto config = SimulatorConfig{
.drop_percent = 0,
.perform_timeouts = true,
.scramble_messages = true,
.rng_seed = 0,
};
std::pair<SimulatorStats, LatencyHistogramSummaries> RunWorkload(SimulatorConfig &config) {
auto simulator = Simulator(config);
auto cli_addr = Address::TestAddress(1);
@ -72,21 +75,47 @@ int main() {
// send request
CounterRequest cli_req;
cli_req.proposal = i;
spdlog::info("[CLIENT] calling Request");
auto res_f = cli_io.Request<CounterRequest, CounterResponse>(srv_addr, cli_req);
spdlog::info("[CLIENT] calling Wait");
auto res_rez = std::move(res_f).Wait();
spdlog::info("[CLIENT] Wait returned");
if (!res_rez.HasError()) {
std::cout << "[CLIENT] Got a valid response" << std::endl;
spdlog::info("[CLIENT] Got a valid response");
auto env = res_rez.GetValue();
MG_ASSERT(env.message.highest_seen == i);
std::cout << "response latency: " << env.response_latency.count() << " microseconds" << std::endl;
spdlog::info("response latency: {} microseconds", env.response_latency.count());
} else {
std::cout << "[CLIENT] Got an error" << std::endl;
spdlog::info("[CLIENT] Got an error");
}
}
using memgraph::utils::print_helpers::operator<<;
std::cout << "response latencies: " << cli_io.ResponseLatencies() << std::endl;
simulator.ShutDown();
return std::make_pair(simulator.Stats(), cli_io.ResponseLatencies());
}
int main() {
spdlog::cfg::load_env_levels();
auto config = SimulatorConfig{
.drop_percent = 0,
.perform_timeouts = true,
.scramble_messages = true,
.rng_seed = 0,
};
auto [sim_stats_1, latency_stats_1] = RunWorkload(config);
auto [sim_stats_2, latency_stats_2] = RunWorkload(config);
if (sim_stats_1 != sim_stats_2 || latency_stats_1 != latency_stats_2) {
spdlog::error("simulator stats diverged across runs");
spdlog::error("run 1 simulator stats: {}", sim_stats_1);
spdlog::error("run 2 simulator stats: {}", sim_stats_2);
spdlog::error("run 1 latency:\n{}", latency_stats_1.SummaryTable());
spdlog::error("run 2 latency:\n{}", latency_stats_2.SummaryTable());
std::terminate();
}
return 0;
}

20
tests/simulation/cluster_property_test.cpp
View File

@ -33,21 +33,31 @@ using io::Time;
using io::simulator::SimulatorConfig;
using storage::v3::kMaximumCronInterval;
RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, NonEmptyOpVec ops)) {
// TODO(tyler) set abort_time to something more restrictive than Time::max()
RC_GTEST_PROP(RandomClusterConfig, HappyPath, (ClusterConfig cluster_config, NonEmptyOpVec ops, uint64_t rng_seed)) {
spdlog::cfg::load_env_levels();
SimulatorConfig sim_config{
.drop_percent = 0,
.perform_timeouts = false,
.scramble_messages = true,
.rng_seed = 0,
.rng_seed = rng_seed,
.start_time = Time::min(),
// TODO(tyler) set abort_time to something more restrictive than Time::max()
.abort_time = Time::max(),
};
RunClusterSimulation(sim_config, cluster_config, ops.ops);
auto [sim_stats_1, latency_stats_1] = RunClusterSimulation(sim_config, cluster_config, ops.ops);
auto [sim_stats_2, latency_stats_2] = RunClusterSimulation(sim_config, cluster_config, ops.ops);
if (latency_stats_1 != latency_stats_2) {
spdlog::error("simulator stats diverged across runs");
spdlog::error("run 1 simulator stats: {}", sim_stats_1);
spdlog::error("run 2 simulator stats: {}", sim_stats_2);
spdlog::error("run 1 latency:\n{}", latency_stats_1.SummaryTable());
spdlog::error("run 2 latency:\n{}", latency_stats_2.SummaryTable());
RC_ASSERT(latency_stats_1 == latency_stats_2);
RC_ASSERT(sim_stats_1 == sim_stats_2);
}
}
} // namespace memgraph::tests::simulation

69
tests/simulation/raft.cpp
View File

@ -14,11 +14,15 @@
#include <iostream>
#include <map>
#include <optional>
#include <random>
#include <set>
#include <thread>
#include <vector>
#include <spdlog/cfg/env.h>
#include "io/address.hpp"
#include "io/message_histogram_collector.hpp"
#include "io/rsm/raft.hpp"
#include "io/rsm/rsm_client.hpp"
#include "io/simulator/simulator.hpp"
@ -27,6 +31,7 @@
using memgraph::io::Address;
using memgraph::io::Duration;
using memgraph::io::Io;
using memgraph::io::LatencyHistogramSummaries;
using memgraph::io::ResponseEnvelope;
using memgraph::io::ResponseFuture;
using memgraph::io::ResponseResult;
@ -123,16 +128,7 @@ void RunRaft(Raft<IoImpl, TestState, CasRequest, CasResponse, GetRequest, GetRes
server.Run();
}
void RunSimulation() {
SimulatorConfig config{
.drop_percent = 5,
.perform_timeouts = true,
.scramble_messages = true,
.rng_seed = 0,
.start_time = Time::min() + std::chrono::microseconds{256 * 1024},
.abort_time = Time::max(),
};
std::pair<SimulatorStats, LatencyHistogramSummaries> RunSimulation(SimulatorConfig &config) {
auto simulator = Simulator(config);
auto cli_addr = Address::TestAddress(1);
@ -189,6 +185,7 @@ void RunSimulation() {
auto write_cas_response_result = client.SendWriteRequest(cas_req);
if (write_cas_response_result.HasError()) {
spdlog::debug("timed out");
// timed out
continue;
}
@ -229,6 +226,10 @@ void RunSimulation() {
simulator.ShutDown();
srv_thread_1.join();
srv_thread_2.join();
srv_thread_3.join();
SimulatorStats stats = simulator.Stats();
spdlog::info("total messages: {}", stats.total_messages);
@ -240,21 +241,51 @@ void RunSimulation() {
spdlog::info("========================== SUCCESS :) ==========================");
/*
this is implicit in jthread's dtor
srv_thread_1.join();
srv_thread_2.join();
srv_thread_3.join();
*/
return std::make_pair(simulator.Stats(), cli_io.ResponseLatencies());
}
void RunWithSeed(uint64_t seed) {
SimulatorConfig config{
.drop_percent = 5,
.perform_timeouts = true,
.scramble_messages = true,
.rng_seed = seed,
.start_time = Time::min() + std::chrono::microseconds{256 * 1024},
.abort_time = Time::min() + std::chrono::seconds{3600},
};
spdlog::error("========================== NEW SIMULATION, replay with RunWithSeed({}) ==========================",
seed);
spdlog::info("\tTime\t\tTerm\tPort\tRole\t\tMessage\n");
auto [sim_stats_1, latency_stats_1] = RunSimulation(config);
spdlog::info("========================== NEW SIMULATION, replay with RunWithSeed({}) ==========================",
seed);
spdlog::info("\tTime\t\tTerm\tPort\tRole\t\tMessage\n");
auto [sim_stats_2, latency_stats_2] = RunSimulation(config);
if (sim_stats_1 != sim_stats_2 || latency_stats_1 != latency_stats_2) {
spdlog::error("simulator stats diverged across runs for test rng_seed: {}", seed);
spdlog::error("run 1 simulator stats: {}", sim_stats_1);
spdlog::error("run 2 simulator stats: {}", sim_stats_2);
spdlog::error("run 1 latency:\n{}", latency_stats_1.SummaryTable());
spdlog::error("run 2 latency:\n{}", latency_stats_2.SummaryTable());
std::terminate();
}
}
int main() {
spdlog::cfg::load_env_levels();
std::random_device random_device;
std::mt19937 generator(random_device());
std::uniform_int_distribution<> distribution;
int n_tests = 50;
for (int i = 0; i < n_tests; i++) {
spdlog::info("========================== NEW SIMULATION {} ==========================", i);
spdlog::info("\tTime\t\tTerm\tPort\tRole\t\tMessage\n");
RunSimulation();
uint64_t seed = distribution(generator);
RunWithSeed(seed);
}
spdlog::info("passed {} tests!", n_tests);

7
tests/simulation/shard_rsm.cpp
View File

@ -1115,11 +1115,12 @@ int TestMessages() {
ConcreteShardRsm shard_server3(std::move(shard_server_io_3), address_for_3, ShardRsm(std::move(shard_ptr3)));
auto server_thread1 = std::jthread([&shard_server1]() { shard_server1.Run(); });
auto server_thread2 = std::jthread([&shard_server2]() { shard_server2.Run(); });
auto server_thread3 = std::jthread([&shard_server3]() { shard_server3.Run(); });
simulator.IncrementServerCountAndWaitForQuiescentState(shard_server_1_address);
auto server_thread2 = std::jthread([&shard_server2]() { shard_server2.Run(); });
simulator.IncrementServerCountAndWaitForQuiescentState(shard_server_2_address);
auto server_thread3 = std::jthread([&shard_server3]() { shard_server3.Run(); });
simulator.IncrementServerCountAndWaitForQuiescentState(shard_server_3_address);
std::cout << "Beginning test after servers have become quiescent." << std::endl;

23
tests/simulation/test_cluster.hpp
View File

@ -24,6 +24,7 @@
#include "coordinator/shard_map.hpp"
#include "generated_operations.hpp"
#include "io/address.hpp"
#include "io/message_histogram_collector.hpp"
#include "io/simulator/simulator.hpp"
#include "io/simulator/simulator_config.hpp"
#include "io/simulator/simulator_transport.hpp"
@ -57,6 +58,7 @@ using io::simulator::SimulatorStats;
using io::simulator::SimulatorTransport;
using machine_manager::MachineConfig;
using machine_manager::MachineManager;
using memgraph::io::LatencyHistogramSummaries;
using msgs::ReadRequests;
using msgs::ReadResponses;
using msgs::WriteRequests;
@ -75,6 +77,8 @@ MachineManager<SimulatorTransport> MkMm(Simulator &simulator, std::vector<Addres
.is_coordinator = true,
.listen_ip = addr.last_known_ip,
.listen_port = addr.last_known_port,
.shard_worker_threads = 4,
.sync_message_handling = true,
};
Io<SimulatorTransport> io = simulator.Register(addr);
@ -210,17 +214,19 @@ struct DetachIfDropped {
}
};
void RunClusterSimulation(const SimulatorConfig &sim_config, const ClusterConfig &cluster_config,
const std::vector<Op> &ops) {
std::pair<SimulatorStats, LatencyHistogramSummaries> RunClusterSimulation(const SimulatorConfig &sim_config,
const ClusterConfig &cluster_config,
const std::vector<Op> &ops) {
spdlog::info("========================== NEW SIMULATION ==========================");
auto simulator = Simulator(sim_config);
auto cli_addr = Address::TestAddress(1);
auto machine_1_addr = cli_addr.ForkUniqueAddress();
auto machine_1_addr = Address::TestAddress(1);
auto cli_addr = Address::TestAddress(2);
auto cli_addr_2 = Address::TestAddress(3);
Io<SimulatorTransport> cli_io = simulator.Register(cli_addr);
Io<SimulatorTransport> cli_io_2 = simulator.Register(Address::TestAddress(2));
Io<SimulatorTransport> cli_io_2 = simulator.Register(cli_addr_2);
auto coordinator_addresses = std::vector{
machine_1_addr,
@ -232,6 +238,7 @@ void RunClusterSimulation(const SimulatorConfig &sim_config, const ClusterConfig
Address coordinator_address = mm_1.CoordinatorAddress();
auto mm_thread_1 = std::jthread(RunMachine, std::move(mm_1));
simulator.IncrementServerCountAndWaitForQuiescentState(machine_1_addr);
auto detach_on_error = DetachIfDropped{.handle = mm_thread_1};
@ -257,6 +264,8 @@ void RunClusterSimulation(const SimulatorConfig &sim_config, const ClusterConfig
simulator.ShutDown();
mm_thread_1.join();
SimulatorStats stats = simulator.Stats();
spdlog::info("total messages: {}", stats.total_messages);
@ -268,10 +277,8 @@ void RunClusterSimulation(const SimulatorConfig &sim_config, const ClusterConfig
auto histo = cli_io_2.ResponseLatencies();
using memgraph::utils::print_helpers::operator<<;
std::cout << "response latencies: " << histo << std::endl;
spdlog::info("========================== SUCCESS :) ==========================");
return std::make_pair(stats, histo);
}
} // namespace memgraph::tests::simulation