perf_fast_forward is used to detect performance regressions. The two
main metrics used for this are fargments per second and the number of
the IO operations. The former is a median of a several runs, but the
latter is just the actual number of asynchronous IO operations performed
in the run that happened to be picked as a median frag/s-wise. There's
no always a direct correlation between frag/s and aio and the latter can
vary which makes the latter hard to compare.
In order to make this easier a new metric was introduced: "average aio"
which reports the average number of asynchronous IO operations performed
in a run. This should produce much more stable results and therefore
make the comparison more meaningful.
Message-Id: <20190430134401.19238-1-pdziepak@scylladb.com>
Rather than {std::experimental,boost,seastar::compat}::filesystem
On Sat, 2019-03-23 at 01:44 +0200, Avi Kivity wrote:
> The intent for seastar::compat was to allow the application to choose
> the C++ dialect and have seastar follow, rather than have seastar choose
> the types and have the application follow (as in your patch).
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
"
Variable length integers are used are used extensively by SSTables mc
format. The current deserialisation routine is quite naive in a way that
it reads each byte separately. Since, those vints usually appear inside
much larger buffers, we optimise for such cases, read 8-bytes at once
and then mask out the unneeded parts (as well as fix their order because
big-endian).
Tests: unit(dev).
perf_vint (average time per element when deserializing 1000 vints):
before:
vint.deserialize 69442000 14.400ns 0.000ns 14.399ns 14.400ns
after:
vint.deserialize 241502000 4.140ns 0.000ns 4.140ns 4.140ns
perf_fast_forward (data on /tmp):
large-partition-single-key-slice on dataset large-part-ds1:
before:
range time (s) iterations frags frag/s mad f/s max f/s min f/s aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu
-> [0, 1] 0.000278 8792 2 7190 119 7367 1960 3 104 2 0 0 1 1 0 0 1 100.0%
-> [1, 100) 0.000344 96 99 288100 4335 307689 193809 2 108 2 0 0 1 1 0 0 1 100.0%
-> (100, 200] 0.000339 13254 100 295263 2824 301734 222725 2 108 2 0 0 1 1 0 0 1 100.0%
after:
range time (s) iterations frags frag/s mad f/s max f/s min f/s aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu
-> [0, 1] 0.000236 10001 2 8461 59 8718 2261 3 104 2 0 0 1 1 0 0 1 100.0%
-> [1, 100) 0.000285 89 99 347500 2441 355826 215745 2 108 2 0 0 1 1 0 0 1 100.0%
-> (100, 200] 0.000293 14369 100 341302 1512 350123 222049 2 108 2 0 0 1 1 0 0 1 100.0%
"
* tag 'optimise-vint/v2' of https://github.com/pdziepak/scylla:
sstable: pass full length of buffer to vint deserialiser
vint: optimise deserialisation routine
vint: drop deserialize_type structure
tests/vint: reduce test dependencies
tests/perf: add performance test for vint serialisation
perf_fast_forward with flag --dump-all-results reports the results of
every test iteration that was executed. This patch introduces a python
script that can analyse those results (in json format) and present them
in a more human-friendly way.
For now, the only option is to plot histograms of selected statistics.
perf_fast_forward runs each test case multiple times and reports a
summary of those results (median, min, max, and median absolute
deviation).
While very convenient the summary may hide some important information
(e.g. the distribution of the results). This patch adds an option to
report results of every single executed iteration.
Several of the test cases in perf_fast_forward do not need
partition-level fast-forwarding. However, since the defaults are used to
construct most of the readers the fast-forwarding is enabled regardless.
This showed an apparent regression in the perf_fast_forward results
after adcb3ec20c ("row_cache: read is not
single-partition if inter-partition forwarding is enabled") which
disabled an optimisation that was invalid when partition-level
fast-forwarind was requested.
This patch ensures that all single-partition reads that do not need
partition-level fast-forwarding keep it disabled.
In preparation for providing a default large_data_handler in
a test-standard way.
buffer_size parameter reordered and now has a default value
same as make_sstable()'s.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
"
This is a first step in fixing #3988.
"
* 'espindola/large-row-warn-only-v4' of https://github.com/espindola/scylla:
Rename large_partition_handler
Print a warning if a row is too large
Remove defaut parameter value
Rename _threshold_bytes to _partition_threshold_bytes
keys: add schema-aware printing for clustering_key_prefix
Otherwise read test results for subsequent datasets will override each other.
Also, rename population test case to not include dataset name, which
is now redundant.
Message-Id: <1547822942-9690-1-git-send-email-tgrabiec@scylladb.com>
Replace stdx::optional and stdx::string_view with the C++ std
counterparts.
Some instances of boost::variant were also replaced with std::variant,
namely those that called seastar::visit.
Scylla now requires GCC 8 to compile.
Signed-off-by: Duarte Nunes <duarte@scylladb.com>
Message-Id: <20190108111141.5369-1-duarte@scylladb.com>
The results vector should be populated vertically, not horizontally.
Responsible for assertion failure with --cache-enabled:
void result_collector::add(test_result_vector): Assertion `rs.size() == results.size()' failed.
Introduced in 3fc78a25bf.
Message-Id: <1544105835-24530-2-git-send-email-tgrabiec@scylladb.com>
Otherwise errors cannot be made sense of, since error are reported
always to stdout. Without test output we don't know what they're
referring to.
This change makes the output always go to stdout, in addition to other
reportes, if any.
Message-Id: <1544020084-16492-1-git-send-email-tgrabiec@scylladb.com>
"
zlib's crc32_combine() is not very efficient. It is faster to re-combine
the buffer using crc32(). It's still substantial amount of work which
could be avoided.
This patch introduces a fast implementation of crc32_combine() which
uses a different algorithm than zlib. It also utilizes intrinsics for
carry-less multiplication instruction to perform the computation faster.
The details of the algorithm can be found in code comments.
Performance results using perf_checksum and second buffer of length 64 KiB:
zlib CRC32 combine: 38'851 ns
libdeflate CRC32: 4'797 ns
fast_crc32_combine(): 11 ns
So the new implementation is 3500x faster than zlib's, and 417x faster than
re-checksumming the buffer using libdeflate.
Tested on i7-5960X CPU @ 3.00GHz
Performance was also evaluated using sstable writer benchmark:
perf_fast_forward --populate --sstable-format=mc --data-directory /tmp/perf-mc \
--value-size=10000 --rows 1000000 --datasets small-part
It yielded 9% improvement in median frag/s (129'055 vs 117'977).
Refs #3874
"
* tag 'fast-crc32-combine-v2' of github.com:tgrabiec/scylla:
tests: perf_checksum: Test fast_crc32_combine()
tests: Rename libdeflate_test to checksum_utils_test
tests: libdeflate: Add more tests for checksum_combine()
tests: libdeflate: Check both libdeflate and default checksummers
sstables: Use fast_crc_combine() in the default checksummer
utils/gz: Add fast implementation of crc32_combine()
utils/gz: Add pre-computed polynomials
utils/gz: Import Barett reduction implementation from libdeflate
utils: Extract clmul() from crc.hh
* tag 'perf-ffwd-dataset-population-v2' of github.com:tgrabiec/scylla:
tests: perf_fast_forward: Measure performance of dataset population
tests: perf_fast_forward: Record the dataset on which test case was run
tests: perf_fast_forward: Introduce the concept of a dataset
tests: perf_fast_forward: Introduce make_compaction_disabling_guard()
tests: perf_fast_forward: Initialize output manager before population
tests: perf_fast_forward: Handle empty test parameter set
tests: perf_fast_forward: Extract json_output_writer::write_common_test_group()
tests: perf_fast_forward: Factor out access to cfg to a single place per function
tests: perf_fast_forward: Extract result_collector
tests: perf_fast_forward: Take writes into account in AIO statistics
tests: perf_fast_forward: Reorder members
tests: perf_fast_forward: Add --sstable-format command line option
Currently, compaction only replace input sstables at end of compaction,
meaning compaction must be finished for all the space of those sstables
to be released.
What we can do instead is to delete earlier some input sstable under
some conditions:
1) SStable data should be committed to a new, sealed output sstable,
meaning it's exhausted.
2) Exhausted sstable mustn't overlap with a non-exhausted sstable
because a tombstone in the exhausted could have been purged and the
shadowed data in non-exhausted could be ressurected if system
crashes.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
A dataset represents a table with data, populated in certain way, with
certain characteristics of the schema and data.
Before this change, datasets were implicitly defined, with population
hard-coded inside the populate() function.
This change gathers logic related to datasets into classes, in order to:
- make it easier to define new datasets.
- be able to measure performance of dataset population in a
standardized way.
- being able to express constraints on datasets imposed by different
test cases. Test cases are matched with possible datasets based
on the abstract interface they accept (e.g. clustered_ds,
multipartition_ds), and which must be implemented by a compatible
dataset. To facilitate this matching, test function is now wrapped
into a dataset_acceptor object, with an automatically-generated can_run()
virtual method, deduced by make_test_fn().
- be able to select tests to run based on the dataset name.
Only tests which are compatible with that dataset will be run.
Extracts the result collection and reporting logic out of
run_test_case(). Will be needed in population tests, for which we
don't want the looping logic.
* seastar d59fcef...b924495 (2):
> build: Fix protobuf generation rules
> Merge "Restructure files" from Jesse
Includes fixup patch from Jesse:
"
Update Seastar `#include`s to reflect restructure
All Seastar header files are now prefixed with "seastar" and the
configure script reflects the new locations of files.
Signed-off-by: Jesse Haber-Kucharsky <jhaberku@scylladb.com>
Message-Id: <5d22d964a7735696fb6bb7606ed88f35dde31413.1542731639.git.jhaberku@scylladb.com>
"
sprint() recently became more strict, throwing on sprint("%s", 5). Replace
with the more modern format().
Mechanically converted with https://github.com/avikivity/unsprint.
sprint() returns std::string(), but the new format() returns an sstring. Usually
an sstring is wanted but in this case an sstring will fail as it is added to
an std::string.
Fix the failure (after spring->format conversion) by converting to an std::string.
