Summary:
Context:
An extra IO_USER priority in rate limiter allows users to optionally charge WAL writes / SST reads to rate limiter at this priority level, which then has higher priority than IO_HIGH and IO_LOW. With an extra IO_USER priority, it allows users to better specify the relative urgency/importance among different requests in rate limiter. As a consequence, IO resource management can better prioritize and limit resource based on user's need.
The IO_USER is implemented as superior priority in GenericRateLimiter, in the sense that its request queue will always be iterated first without being constrained to fairness. The reason is that the notion of fairness is only meaningful in helping lower priorities in background IO (i.e, IO_HIGH/MID/LOW) to gain some fair chance to run so that it does not block foreground IO (i.e, the ones that are charged at the level of IO_USER). As we can see, the ultimate goal here is to not blocking foreground IO at IO_USER level, which justifies the superiority of IO_USER.
Similar benefits exist for IO_MID priority.
- Rewrote the logic of deciding the order of iterating request queues of high/low priorities to include the extra user/mid priority w/o affecting the existing behavior (see PR's [comment](https://github.com/facebook/rocksdb/pull/8595/files#r678749331))
- Included the request queue of user-pri/mid-pri in the code path of next-leader-candidate signaling and GenericRateLimiter's destructor
- Included the extra user/mid-pri in bookkeeping data structures: total_bytes_through_ and total_requests_
- Re-written the previous impl of explicitly iterating priorities with a loop from Env::IO_LOW to Env::IO_TOTAL
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8595
Test Plan:
- passed existing rate_limiter_test.cc
- passed added unit tests in rate_limiter_test.cc
- run performance test to verify performance with only high/low requests is not affected by this change
- Set-up command:
`TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom --duration=5 --compression_type=none --num=100000000 --disable_auto_compactions=true --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1))`
- Test command:
`TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=overwrite --use_existing_db=true --disable_wal=true --duration=30 --compression_type=none --num=100000000 --write_buffer_size=1048576 --writable_file_max_buffer_size=65536 --target_file_size_base=1048576 --max_bytes_for_level_base=4194304 --level0_slowdown_writes_trigger=$(((1 << 31) - 1)) --level0_stop_writes_trigger=$(((1 << 31) - 1)) --statistics=true --rate_limiter_bytes_per_sec=1048576 --rate_limiter_refill_period_us=1000 --threads=32 |& grep -E '(flush|compact)\.write\.bytes'`
- Before (on branch upstream/master):
`rocksdb.compact.write.bytes COUNT : 4014162`
`rocksdb.flush.write.bytes COUNT : 26715832`
rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.66
- After (on branch rate_limiter_user_pri):
`rocksdb.compact.write.bytes COUNT : 3807822`
`rocksdb.flush.write.bytes COUNT : 26098659`
rocksdb.flush.write.bytes/rocksdb.compact.write.bytes ~= 6.85
Reviewed By: ajkr
Differential Revision: D30577783
Pulled By: hx235
fbshipit-source-id: 0881f2705ffd13ecd331256bde7e8ec874a353f4
Summary:
`GenericRateLimiter` slow path handles requests that cannot be satisfied
immediately. Such requests enter a queue, and their thread stays in `Request()`
until they are granted or the rate limiter is stopped. These threads are
responsible for unblocking themselves. The work to do so is split into two main
duties.
(1) Waiting for the next refill time.
(2) Refilling the bytes and granting requests.
Prior to this PR, the slow path logic involved a leader election algorithm to
pick one thread to perform (1) followed by (2). It elected the thread whose
request was at the front of the highest priority non-empty queue since that
request was most likely to be granted. This algorithm was efficient in terms of
reducing intermediate wakeups, which is a thread waking up only to resume
waiting after finding its request is not granted. However, the conceptual
complexity of this algorithm was too high. It took me a long time to draw a
timeline to understand how it works for just one edge case yet there were so
many.
This PR drops the leader election to reduce conceptual complexity. Now, the two
duties can be performed by whichever thread acquires the lock first. The risk
of this change is increasing the number of intermediate wakeups, however, we
took steps to mitigate that.
- `wait_until_refill_pending_` flag ensures only one thread performs (1). This\
prevents the thundering herd problem at the next refill time. The remaining\
threads wait on their condition variable with an unbounded duration -- thus we\
must remember to notify them to ensure forward progress.
- (1) is typically done by a thread at the front of a queue. This is trivial\
when the queues are initially empty as the first choice that arrives must be\
the only entry in its queue. When queues are initially non-empty, we achieve\
this by having (2) notify a thread at the front of a queue (preferring higher\
priority) to perform the next duty.
- We do not require any additional wakeup for (2). Typically it will just be\
done by the thread that finished (1).
Combined, the second and third bullet points above suggest the refill/granting
will typically be done by a request at the front of its queue. This is
important because one wakeup is saved when a granted request happens to be in an
already running thread.
Note there are a few cases that still lead to intermediate wakeup, however. The
first two are existing issues that also apply to the old algorithm, however, the
third (including both subpoints) is new.
- No request may be granted (only possible when rate limit dynamically\
decreases).
- Requests from a different queue may be granted.
- (2) may be run by a non-front request thread causing it to not be granted even\
if some requests in that same queue are granted. It can happen for a couple\
(unlikely) reasons.
- A new request may sneak in and grab the lock at the refill time, before the\
thread finishing (1) can wake up and grab it.
- A new request may sneak in and grab the lock and execute (1) before (2)'s\
chosen candidate can wake up and grab the lock. Then that non-front request\
thread performing (1) can carry over to perform (2).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8602
Test Plan:
- Use existing tests. The edge cases listed in the comment are all performance\
related; I could not really think of any related to correctness. The logic\
looks the same whether a thread wakes up/finishes its work early/on-time/late,\
or whether the thread is chosen vs. "steals" the work.
- Verified write throughput and CPU overhead are basically the same with and\
without this change, even in a rate limiter heavy workload:
Test command:
```
$ rm -rf /dev/shm/dbbench/ && TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -benchmarks=fillrandom -num_multi_db=64 -num_low_pri_threads=64 -num_high_pri_threads=64 -write_buffer_size=262144 -target_file_size_base=262144 -max_bytes_for_level_base=1048576 -rate_limiter_bytes_per_sec=16777216 -key_size=24 -value_size=1000 -num=10000 -compression_type=none -rate_limiter_refill_period_us=1000
```
Results before this PR:
```
fillrandom : 108.463 micros/op 9219 ops/sec; 9.0 MB/s
7.40user 8.84system 1:26.20elapsed 18%CPU (0avgtext+0avgdata 256140maxresident)k
```
Results after this PR:
```
fillrandom : 108.108 micros/op 9250 ops/sec; 9.0 MB/s
7.45user 8.23system 1:26.68elapsed 18%CPU (0avgtext+0avgdata 255688maxresident)k
```
Reviewed By: hx235
Differential Revision: D30048013
Pulled By: ajkr
fbshipit-source-id: 6741bba9d9dfbccab359806d725105817fef818b
Summary:
Context:
As need for new feature of resource management using RocksDB's rate limiter like [https://github.com/facebook/rocksdb/issues/8595](https://github.com/facebook/rocksdb/pull/8595) arises, it is about time to re-learn our rate limiter and make this learning process easier for others by improving its readability. The comment/assertion/one extra else-branch are added based on my best understanding toward the rate_limiter.cc and rate_limiter_test.cc up to date after giving it a hard read.
- Add code comments/assertion/one extra else-branch (that is not affecting existing behavior, see PR comment) to describe how leader-election works under multi-thread settings in GenericRateLimiter::Request()
- Add code comments to describe a non-obvious trick during clean-up of rate limiter destructor
- Add code comments to explain more about the starvation being fixed in GenericRateLimiter::Refill() through partial byte-granting
- Add code comments to the rate limiter's setup in a complicated unit test in rate_limiter_test
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8596
Test Plan: - passed existing rate_limiter_test.cc
Reviewed By: ajkr
Differential Revision: D29982590
Pulled By: hx235
fbshipit-source-id: c3592986bb5b0c90d8229fe44f425251ec7e8a0a
Summary:
For performance purposes, the lower level routines were changed to use a SystemClock* instead of a std::shared_ptr<SystemClock>. The shared ptr has some performance degradation on certain hardware classes.
For most of the system, there is no risk of the pointer being deleted/invalid because the shared_ptr will be stored elsewhere. For example, the ImmutableDBOptions stores the Env which has a std::shared_ptr<SystemClock> in it. The SystemClock* within the ImmutableDBOptions is essentially a "short cut" to gain access to this constant resource.
There were a few classes (PeriodicWorkScheduler?) where the "short cut" property did not hold. In those cases, the shared pointer was preserved.
Using db_bench readrandom perf_level=3 on my EC2 box, this change performed as well or better than 6.17:
6.17: readrandom : 28.046 micros/op 854902 ops/sec; 61.3 MB/s (355999 of 355999 found)
6.18: readrandom : 32.615 micros/op 735306 ops/sec; 52.7 MB/s (290999 of 290999 found)
PR: readrandom : 27.500 micros/op 871909 ops/sec; 62.5 MB/s (367999 of 367999 found)
(Note that the times for 6.18 are prior to revert of the SystemClock).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8033
Reviewed By: pdillinger
Differential Revision: D27014563
Pulled By: mrambacher
fbshipit-source-id: ad0459eba03182e454391b5926bf5cdd45657b67
Summary:
Introduces and uses a SystemClock class to RocksDB. This class contains the time-related functions of an Env and these functions can be redirected from the Env to the SystemClock.
Many of the places that used an Env (Timer, PerfStepTimer, RepeatableThread, RateLimiter, WriteController) for time-related functions have been changed to use SystemClock instead. There are likely more places that can be changed, but this is a start to show what can/should be done. Over time it would be nice to migrate most (if not all) of the uses of the time functions from the Env to the SystemClock.
There are several Env classes that implement these functions. Most of these have not been converted yet to SystemClock implementations; that will come in a subsequent PR. It would be good to unify many of the Mock Timer implementations, so that they behave similarly and be tested similarly (some override Sleep, some use a MockSleep, etc).
Additionally, this change will allow new methods to be introduced to the SystemClock (like https://github.com/facebook/rocksdb/issues/7101 WaitFor) in a consistent manner across a smaller number of classes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7858
Reviewed By: pdillinger
Differential Revision: D26006406
Pulled By: mrambacher
fbshipit-source-id: ed10a8abbdab7ff2e23d69d85bd25b3e7e899e90
Summary:
When dynamically linking two binaries together, different builds of RocksDB from two sources might cause errors. To provide a tool for user to solve the problem, the RocksDB namespace is changed to a flag which can be overridden in build time.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6433
Test Plan: Build release, all and jtest. Try to build with ROCKSDB_NAMESPACE with another flag.
Differential Revision: D19977691
fbshipit-source-id: aa7f2d0972e1c31d75339ac48478f34f6cfcfb3e
Summary:
Dynamic adjustment of rate limit according to demand for background I/O. It increases by a factor when limiter is drained too frequently, and decreases by the same factor when limiter is not drained frequently enough. The parameters for this behavior are fixed in `GenericRateLimiter::Tune`. Other changes:
- make rate limiter's `Env*` configurable for testing
- track num drain intervals in RateLimiter so we don't have to rely on stats, which may be shared across different DB instances from the ones that share the RateLimiter.
Closes https://github.com/facebook/rocksdb/pull/2899
Differential Revision: D5858704
Pulled By: ajkr
fbshipit-source-id: cc2bac30f85e7f6fd63655d0a6732ef9ed7403b1
Summary:
Allow users to rate limit background work based on read bytes, written bytes, or sum of read and written bytes. Support these by changing the RateLimiter API, so no additional options were needed.
Closes https://github.com/facebook/rocksdb/pull/2433
Differential Revision: D5216946
Pulled By: ajkr
fbshipit-source-id: aec57a8357dbb4bfde2003261094d786d94f724e
Summary:
It's hard for RocksDB to come up with a good default of delayed write rate. Use rate given by rate limiter if it is availalbe. This provides the I/O order of magnitude.
Closes https://github.com/facebook/rocksdb/pull/2357
Differential Revision: D5115324
Pulled By: siying
fbshipit-source-id: 341065ad2211c981fc804011c0f0e59a50c7e754
Summary:
This is the metric I plan to use for adaptive rate limiting. The statistics are updated only if the rate limiter is drained by flush or compaction. I believe (but am not certain) that this is the normal case.
The Statistics object is passed in RateLimiter::Request() to avoid requiring changes to client code, which would've been necessary if we passed it in the RateLimiter constructor.
Closes https://github.com/facebook/rocksdb/pull/1946
Differential Revision: D4646489
Pulled By: ajkr
fbshipit-source-id: d8e0161
Summary:
NowMicros() provides non-monotonic time. When wall clock is
synchronized or changed, the non-monotonicity time points will affect write rate
controllers. This patch changes write_controller.cc and rate_limiter.cc to use
monotonic time points.
Closes https://github.com/facebook/rocksdb/pull/1865
Differential Revision: D4561732
Pulled By: siying
fbshipit-source-id: 95ece62
Summary: When rate_bytes_per_sec * refill_period_us_ overflows, the actual limited rate is very low. Handle this case so the rate will be large.
Test Plan: Add a unit test for it.
Reviewers: IslamAbdelRahman, andrewkr
Reviewed By: andrewkr
Subscribers: yiwu, lightmark, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D58929
Summary: If the user specified a small enough value for the rate limiter's bytes per second, the calculation for the number of refill bytes per period could become zero which would effectively cause the server to hang forever.
Test Plan: Existing tests
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56631
Summary: BackupRateLimiter removed and uses replaced with the existing GenericRateLimiter
Test Plan:
make all check
make clean
USE_CLANG=1 make all
make clean
OPT=-DROCKSDB_LITE make release
Reviewers: leveldb, igor
Reviewed By: igor
Subscribers: igor, dhruba
Differential Revision: https://reviews.facebook.net/D46095
Summary: Make RocksDb build and run on Windows to be functionally
complete and performant. All existing test cases run with no
regressions. Performance numbers are in the pull-request.
Test plan: make all of the existing unit tests pass, obtain perf numbers.
Co-authored-by: Praveen Rao praveensinghrao@outlook.com
Co-authored-by: Sherlock Huang baihan.huang@gmail.com
Co-authored-by: Alex Zinoviev alexander.zinoviev@me.com
Co-authored-by: Dmitri Smirnov dmitrism@microsoft.com
Summary: This feature is going to be useful for mongodb+rocksdb. I'll expose it through mongo's API.
Test Plan: added new unit test. also will run TSAN on the new unit test
Reviewers: meyering, sdong
Reviewed By: meyering, sdong
Subscribers: meyering, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D35307
Summary:
User gets undefinied error since the definition is not exposed.
Also re-enable the db test with only upper bound check
Test Plan: db_test, rate_limit_test
Reviewers: igor, yhchiang, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D20403
Summary:
A generic rate limiter that can be shared by threads and rocksdb
instances. Will use this to smooth out write traffic generated by
compaction and flush. This will help us get better p99 behavior on flash
storage.
Test Plan:
unit test output
==== Test RateLimiterTest.Rate
request size [1 - 1023], limit 10 KB/sec, actual rate: 10.374969 KB/sec, elapsed 2002265
request size [1 - 2047], limit 20 KB/sec, actual rate: 20.771242 KB/sec, elapsed 2002139
request size [1 - 4095], limit 40 KB/sec, actual rate: 41.285299 KB/sec, elapsed 2202424
request size [1 - 8191], limit 80 KB/sec, actual rate: 81.371605 KB/sec, elapsed 2402558
request size [1 - 16383], limit 160 KB/sec, actual rate: 162.541268 KB/sec, elapsed 3303500
Reviewers: yhchiang, igor, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D19359
