|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
Push- instead of pull-model for managing Write stalls
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
10 years ago
|
|
|
//
|
|
|
|
#include "db/write_controller.h"
|
|
|
|
|
|
|
|
#include <array>
|
|
|
|
#include <ratio>
|
|
|
|
|
|
|
|
#include "rocksdb/system_clock.h"
|
|
|
|
#include "test_util/testharness.h"
|
Push- instead of pull-model for managing Write stalls
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
10 years ago
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
namespace {
|
|
|
|
class TimeSetClock : public SystemClockWrapper {
|
|
|
|
public:
|
|
|
|
explicit TimeSetClock() : SystemClockWrapper(nullptr) {}
|
|
|
|
const char* Name() const override { return "TimeSetClock"; }
|
|
|
|
uint64_t now_micros_ = 6666;
|
|
|
|
uint64_t NowNanos() override { return now_micros_ * std::milli::den; }
|
|
|
|
};
|
|
|
|
} // anonymous namespace
|
|
|
|
class WriteControllerTest : public testing::Test {
|
|
|
|
public:
|
|
|
|
WriteControllerTest() { clock_ = std::make_shared<TimeSetClock>(); }
|
|
|
|
std::shared_ptr<TimeSetClock> clock_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Make tests easier to read
|
|
|
|
#define MILLION *1000000u
|
|
|
|
#define MB MILLION
|
|
|
|
#define MBPS MILLION
|
|
|
|
#define SECS MILLION // in microseconds
|
|
|
|
|
|
|
|
TEST_F(WriteControllerTest, BasicAPI) {
|
|
|
|
WriteController controller(40 MBPS); // also set max delayed rate
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 40 MBPS);
|
|
|
|
EXPECT_FALSE(controller.IsStopped());
|
|
|
|
EXPECT_FALSE(controller.NeedsDelay());
|
|
|
|
EXPECT_EQ(0, controller.GetDelay(clock_.get(), 100 MB));
|
|
|
|
|
|
|
|
// set, get
|
|
|
|
controller.set_delayed_write_rate(20 MBPS);
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 20 MBPS);
|
|
|
|
EXPECT_FALSE(controller.IsStopped());
|
|
|
|
EXPECT_FALSE(controller.NeedsDelay());
|
|
|
|
EXPECT_EQ(0, controller.GetDelay(clock_.get(), 100 MB));
|
|
|
|
|
|
|
|
{
|
|
|
|
// set with token, get
|
|
|
|
auto delay_token_0 = controller.GetDelayToken(10 MBPS);
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 10 MBPS);
|
|
|
|
EXPECT_FALSE(controller.IsStopped());
|
|
|
|
EXPECT_TRUE(controller.NeedsDelay());
|
|
|
|
// test with delay
|
|
|
|
EXPECT_EQ(2 SECS, controller.GetDelay(clock_.get(), 20 MB));
|
|
|
|
clock_->now_micros_ += 2 SECS; // pay the "debt"
|
|
|
|
|
|
|
|
auto delay_token_1 = controller.GetDelayToken(2 MBPS);
|
|
|
|
EXPECT_EQ(10 SECS, controller.GetDelay(clock_.get(), 20 MB));
|
|
|
|
clock_->now_micros_ += 10 SECS; // pay the "debt"
|
|
|
|
|
|
|
|
auto delay_token_2 = controller.GetDelayToken(1 MBPS);
|
|
|
|
EXPECT_EQ(20 SECS, controller.GetDelay(clock_.get(), 20 MB));
|
|
|
|
clock_->now_micros_ += 20 SECS; // pay the "debt"
|
|
|
|
|
|
|
|
auto delay_token_3 = controller.GetDelayToken(20 MBPS);
|
|
|
|
EXPECT_EQ(1 SECS, controller.GetDelay(clock_.get(), 20 MB));
|
|
|
|
clock_->now_micros_ += 1 SECS; // pay the "debt"
|
|
|
|
|
|
|
|
// 60M is more than the max rate of 40M. Max rate will be used.
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 20 MBPS);
|
|
|
|
auto delay_token_4 =
|
|
|
|
controller.GetDelayToken(controller.delayed_write_rate() * 3);
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 40 MBPS);
|
|
|
|
EXPECT_EQ(static_cast<uint64_t>(0.5 SECS),
|
|
|
|
controller.GetDelay(clock_.get(), 20 MB));
|
|
|
|
|
|
|
|
EXPECT_FALSE(controller.IsStopped());
|
|
|
|
EXPECT_TRUE(controller.NeedsDelay());
|
|
|
|
|
|
|
|
// Test stop tokens
|
|
|
|
{
|
|
|
|
auto stop_token_1 = controller.GetStopToken();
|
|
|
|
EXPECT_TRUE(controller.IsStopped());
|
|
|
|
EXPECT_EQ(0, controller.GetDelay(clock_.get(), 100 MB));
|
|
|
|
{
|
|
|
|
auto stop_token_2 = controller.GetStopToken();
|
|
|
|
EXPECT_TRUE(controller.IsStopped());
|
|
|
|
EXPECT_EQ(0, controller.GetDelay(clock_.get(), 100 MB));
|
|
|
|
}
|
|
|
|
EXPECT_TRUE(controller.IsStopped());
|
|
|
|
EXPECT_EQ(0, controller.GetDelay(clock_.get(), 100 MB));
|
|
|
|
}
|
|
|
|
// Stop tokens released
|
|
|
|
EXPECT_FALSE(controller.IsStopped());
|
|
|
|
EXPECT_TRUE(controller.NeedsDelay());
|
|
|
|
EXPECT_EQ(controller.delayed_write_rate(), 40 MBPS);
|
|
|
|
// pay the previous "debt"
|
|
|
|
clock_->now_micros_ += static_cast<uint64_t>(0.5 SECS);
|
|
|
|
EXPECT_EQ(1 SECS, controller.GetDelay(clock_.get(), 40 MB));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delay tokens released
|
|
|
|
EXPECT_FALSE(controller.NeedsDelay());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(WriteControllerTest, StartFilled) {
|
|
|
|
WriteController controller(10 MBPS);
|
|
|
|
|
|
|
|
// Attempt to write two things that combined would be allowed within
|
|
|
|
// a single refill interval
|
When slowdown is triggered, reduce the write rate
Summary: It's usually hard for users to set a value of options.delayed_write_rate. With this diff, after slowdown condition triggers, we greedily reduce write rate if estimated pending compaction bytes increase. If estimated compaction pending bytes drop, we increase the write rate.
Test Plan:
Add a unit test
Test with db_bench setting:
TEST_TMPDIR=/dev/shm/ ./db_bench --benchmarks=fillrandom -num=10000000 --soft_pending_compaction_bytes_limit=1000000000 --hard_pending_compaction_bytes_limit=3000000000 --delayed_write_rate=100000000
and make sure without the commit, write stop will happen, but with the commit, it will not happen.
Reviewers: igor, anthony, rven, yhchiang, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52131
9 years ago
|
|
|
auto delay_token_0 =
|
|
|
|
controller.GetDelayToken(controller.delayed_write_rate());
|
|
|
|
|
|
|
|
// Verify no delay because write rate has not been exceeded within
|
|
|
|
// refill interval.
|
|
|
|
EXPECT_EQ(0U, controller.GetDelay(clock_.get(), 2000u /*bytes*/));
|
|
|
|
EXPECT_EQ(0U, controller.GetDelay(clock_.get(), 2000u /*bytes*/));
|
|
|
|
|
|
|
|
// Allow refill (kMicrosPerRefill)
|
|
|
|
clock_->now_micros_ += 1000;
|
|
|
|
|
|
|
|
// Again
|
|
|
|
EXPECT_EQ(0U, controller.GetDelay(clock_.get(), 2000u /*bytes*/));
|
|
|
|
EXPECT_EQ(0U, controller.GetDelay(clock_.get(), 2000u /*bytes*/));
|
|
|
|
|
|
|
|
// Control: something bigger that would exceed write rate within interval
|
|
|
|
uint64_t delay = controller.GetDelay(clock_.get(), 10 MB);
|
|
|
|
EXPECT_GT(1.0 * delay, 0.999 SECS);
|
|
|
|
EXPECT_LT(1.0 * delay, 1.001 SECS);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST_F(WriteControllerTest, DebtAccumulation) {
|
|
|
|
WriteController controller(10 MBPS);
|
|
|
|
|
|
|
|
std::array<std::unique_ptr<WriteControllerToken>, 10> tokens;
|
|
|
|
|
|
|
|
// Accumulate a time delay debt with no passage of time, like many column
|
|
|
|
// families delaying writes simultaneously. (Old versions of WriteController
|
|
|
|
// would reset the debt on every GetDelayToken.)
|
|
|
|
uint64_t debt = 0;
|
|
|
|
for (unsigned i = 0; i < tokens.size(); ++i) {
|
|
|
|
tokens[i] = controller.GetDelayToken((i + 1u) MBPS);
|
|
|
|
uint64_t delay = controller.GetDelay(clock_.get(), 63 MB);
|
|
|
|
ASSERT_GT(delay, debt);
|
|
|
|
uint64_t incremental = delay - debt;
|
|
|
|
ASSERT_EQ(incremental, (63 SECS) / (i + 1u));
|
|
|
|
debt += incremental;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Pay down the debt
|
|
|
|
clock_->now_micros_ += debt;
|
|
|
|
debt = 0;
|
|
|
|
|
|
|
|
// Now accumulate debt with some passage of time.
|
|
|
|
for (unsigned i = 0; i < tokens.size(); ++i) {
|
|
|
|
// Debt is accumulated in time, not in bytes, so this new write
|
|
|
|
// limit is not applied to prior requested delays, even it they are
|
|
|
|
// in progress.
|
|
|
|
tokens[i] = controller.GetDelayToken((i + 1u) MBPS);
|
|
|
|
uint64_t delay = controller.GetDelay(clock_.get(), 63 MB);
|
|
|
|
ASSERT_GT(delay, debt);
|
|
|
|
uint64_t incremental = delay - debt;
|
|
|
|
ASSERT_EQ(incremental, (63 SECS) / (i + 1u));
|
|
|
|
debt += incremental;
|
|
|
|
uint64_t credit = debt / 2;
|
|
|
|
clock_->now_micros_ += credit;
|
|
|
|
debt -= credit;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Pay down the debt
|
|
|
|
clock_->now_micros_ += debt;
|
|
|
|
debt = 0; // consistent state
|
|
|
|
(void)debt; // appease clang-analyze
|
|
|
|
|
|
|
|
// Verify paid down
|
|
|
|
EXPECT_EQ(0U, controller.GetDelay(clock_.get(), 100u /*small bytes*/));
|
|
|
|
|
|
|
|
// Accumulate another debt, without accounting, and releasing tokens
|
|
|
|
for (unsigned i = 0; i < tokens.size(); ++i) {
|
|
|
|
// Big and small are delayed
|
|
|
|
ASSERT_LT(0U, controller.GetDelay(clock_.get(), 63 MB));
|
|
|
|
ASSERT_LT(0U, controller.GetDelay(clock_.get(), 100u /*small bytes*/));
|
|
|
|
tokens[i].reset();
|
|
|
|
}
|
|
|
|
// All tokens released.
|
|
|
|
// Verify that releasing all tokens pays down debt, even with no time passage.
|
|
|
|
tokens[0] = controller.GetDelayToken(1 MBPS);
|
|
|
|
ASSERT_EQ(0U, controller.GetDelay(clock_.get(), 100u /*small bytes*/));
|
When slowdown is triggered, reduce the write rate
Summary: It's usually hard for users to set a value of options.delayed_write_rate. With this diff, after slowdown condition triggers, we greedily reduce write rate if estimated pending compaction bytes increase. If estimated compaction pending bytes drop, we increase the write rate.
Test Plan:
Add a unit test
Test with db_bench setting:
TEST_TMPDIR=/dev/shm/ ./db_bench --benchmarks=fillrandom -num=10000000 --soft_pending_compaction_bytes_limit=1000000000 --hard_pending_compaction_bytes_limit=3000000000 --delayed_write_rate=100000000
and make sure without the commit, write stop will happen, but with the commit, it will not happen.
Reviewers: igor, anthony, rven, yhchiang, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52131
9 years ago
|
|
|
}
|
|
|
|
|
|
|
|
// This may or may not be a "good" feature, but it's an old feature
|
|
|
|
TEST_F(WriteControllerTest, CreditAccumulation) {
|
|
|
|
WriteController controller(10 MBPS);
|
|
|
|
|
|
|
|
std::array<std::unique_ptr<WriteControllerToken>, 10> tokens;
|
|
|
|
|
|
|
|
// Ensure started
|
|
|
|
tokens[0] = controller.GetDelayToken(1 MBPS);
|
|
|
|
ASSERT_EQ(10 SECS, controller.GetDelay(clock_.get(), 10 MB));
|
|
|
|
clock_->now_micros_ += 10 SECS;
|
|
|
|
|
|
|
|
// Accumulate a credit
|
|
|
|
uint64_t credit = 1000 SECS /* see below: * 1 MB / 1 SEC */;
|
|
|
|
clock_->now_micros_ += credit;
|
|
|
|
|
|
|
|
// Spend some credit (burst of I/O)
|
|
|
|
for (unsigned i = 0; i < tokens.size(); ++i) {
|
|
|
|
tokens[i] = controller.GetDelayToken((i + 1u) MBPS);
|
|
|
|
ASSERT_EQ(0U, controller.GetDelay(clock_.get(), 63 MB));
|
|
|
|
// In WriteController, credit is accumulated in bytes, not in time.
|
|
|
|
// After an "unnecessary" delay, all of our time credit will be
|
|
|
|
// translated to bytes on the next operation, in this case with
|
|
|
|
// setting 1 MBPS. So regardless of the rate at delay time, we just
|
|
|
|
// account for the bytes.
|
|
|
|
credit -= 63 MB;
|
|
|
|
}
|
|
|
|
// Spend remaining credit
|
|
|
|
tokens[0] = controller.GetDelayToken(1 MBPS);
|
|
|
|
ASSERT_EQ(0U, controller.GetDelay(clock_.get(), credit));
|
|
|
|
// Verify
|
|
|
|
ASSERT_EQ(10 SECS, controller.GetDelay(clock_.get(), 10 MB));
|
|
|
|
clock_->now_micros_ += 10 SECS;
|
|
|
|
|
|
|
|
// Accumulate a credit, no accounting
|
|
|
|
clock_->now_micros_ += 1000 SECS;
|
|
|
|
|
|
|
|
// Spend a small amount, releasing tokens
|
|
|
|
for (unsigned i = 0; i < tokens.size(); ++i) {
|
|
|
|
ASSERT_EQ(0U, controller.GetDelay(clock_.get(), 3 MB));
|
|
|
|
tokens[i].reset();
|
|
|
|
}
|
|
|
|
|
|
|
|
// All tokens released.
|
|
|
|
// Verify credit is wiped away on new delay.
|
|
|
|
tokens[0] = controller.GetDelayToken(1 MBPS);
|
|
|
|
ASSERT_EQ(10 SECS, controller.GetDelay(clock_.get(), 10 MB));
|
Push- instead of pull-model for managing Write stalls
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
10 years ago
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|
Push- instead of pull-model for managing Write stalls
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
10 years ago
|
|
|
|
|
|
|
int main(int argc, char** argv) {
|
|
|
|
ROCKSDB_NAMESPACE::port::InstallStackTraceHandler();
|
|
|
|
::testing::InitGoogleTest(&argc, argv);
|
|
|
|
return RUN_ALL_TESTS();
|
|
|
|
}
|