1Sereal::Performance(3)User Contributed Perl DocumentationSereal::Performance(3)
2
6 Sereal::Performance - Getting the most out of the Perl-Sereal
7 implementation
8
10 # This is different from the standard module synopsis in
11 # that it chooses performance over ease-of-use.
12 # Think twice before micro-optimizing your Sereal usage.
13 # Usually, Sereal is a lot faster than most of one's code,
14 # so unless you are doing bulk encoding/decoding, you are
15 # better off optimizing for maintainability.
16 use Sereal qw(sereal_encode_with_object
18 sereal_decode_with_object);
19 my $enc = Sereal::Encoder->new();
20 my $dec = Sereal::Decoder->new();
21 my $big_data_structure = {...};
23 my $srldoc = sereal_encode_with_object($enc, $big_data_structure);
25 my $and_back = sereal_decode_with_object($dec, $srldoc);
27
29 Using Sereal in the way that is optimally performant for your use case
30 can make quite a significant difference in performance. Broadly
31 speaking, there are two classes of tweaks you can do: choosing the
32 right options during encoding (sometimes incurring trade-offs in output
33 size) and calling the Sereal encode/decode functions in the most
34 efficient way.
35 If you are not yet using re-usable Sereal::Encoder and Sereal::Decoder
37 objects, then read no further. By switching from the "encode_sereal"
38 and "decode_sereal" functions to either the OO interface or the
39 advanced functional interface, you will get a noticeable speed boost as
40 encoder and decoder structures can be reused. This is particularly
41 significant for the encoder, which can re-use its output buffer. In
42 some cases, such a warmed-up encoder can avoid most memory allocations.
43 I repeat, if you care about performance, then do not use the
45 "encode_sereal" and "decode_sereal" interface.
46 The exact performance in time and space depends heavily on the data
48 structure to be (de-)serialized. Often there is a trade-off between
49 space and time. If in doubt, do your own testing and most importantly
50 ALWAYS TEST WITH REAL DATA. If you care purely about speed at the
51 expense of output size, you can use the "no_shared_hashkeys" option for
52 a small speed-up, see below. If you need smaller output at the cost of
53 higher CPU load and more memory used during encoding/decoding, try the
54 "dedupe_strings" option and enable Snappy compression.
55 For ready-made comparison scripts, see the author_tools/bench.pl and
57 author_tools/dbench.pl programs that are part of this distribution.
58 Suffice to say that this library is easily competitive in both time and
59 space efficiency with the best alternatives.
60 If switching to the OO interface is not enough, you may consider
62 switching to the advanced functional interface that avoids method
63 lookup overhead, and by inlining as custom Perl OPs, may also avoid
64 some of the Perl function call overhead (Perl 5.14 and up). This
65 additional speed-up is only a constant-offset, avoiding said
66 method/function call, rather than speeding up encoding itself and so
67 will be most significant if you are working with very small data sets.
68 "sereal_encode_with_object" and "sereal_decode_with_object" are
70 optionally exported from the Sereal module (or "Sereal::Encoder" and
71 "Sereal::Decoder" respectively). They work the same as the object-
72 oriented interface except that they are invoked differently:
73 $srl_doc = $encoder->encode($data);
75 becomes
77 $srl_doc = sereal_encode_with_object($encoder, $data);
79 and
81 $data = $decoder->decode($srl_doc);
83 becomes
85 $data = sereal_decode_with_object($decoder, $srl_doc);
87 On Perl versions before 5.14, this will be marginally faster than the
89 OO interface as it avoids method lookup. This should rarely matter. On
90 Perl versions starting from 5.14, the function call to
91 "sereal_encode_with_object" or "sereal_decode_with_object" will also be
92 replaced with a custom Perl OP, thus avoiding most of the function call
93 overhead as well.
94 Tuning the "Sereal::Encoder"
96 Several of the "Sereal::Encoder" options add or remove useful behaviour
97 and some of them come at a runtime performance cost.
98 "no_shared_hashkeys"
100 By default, Sereal will emit a "repetition" marker for hash keys that
101 were already previously encountered. Depending on your data
102 structure, this can save quite a bit of space in the generated
103 document. Consider, for example, encoding an array of many objects of
104 the same class. But it may not save anything if you don't have a lot
105 of repeated hash keys or don't even encode any hashes to begin with.
106 In those cases, you can turn this feature off with the
108 "no_shared_hashkeys" option for a small but measurable speed-up.
109 "dedupe_strings"
111 If set, this option will apply the de-duplication logic to all
112 strings that is only applied to hash keys by default. This can be
113 quite expensive in both memory and performance. The same is true for
114 "aliased_dedupe_strings".
115 "snappy" and "snappy_incr"
117 Enabling Snappy compression can (but doesn't have to) make your
118 Sereal documents significantly smaller. How effective this
119 compression is for you depends entirely on the nature of your data.
120 Snappy compression is designed to be very fast. The additional space
121 savings are very often worth the small overhead.
122 "freeze_callbacks"
124 Using custom Perl "FREEZE" callbacks is very expensive. If enabled,
125 the encoder has to do a method lookup at least once per class of an
126 object being serialized. If a "FREEZE" hook actually exists, calling
127 it will be even more expensive. If you care about ultimate
128 performance, use with care.
129 "sort_keys"
131 This option forces the encoder to always "sort" the entries in a hash
132 by its keys before writing them to the Sereal document. This can be
133 somewhat expensive for large hashes.
134 General Considerations
136 Perl variables (scalars specifically) can, at the same time, hold
137 multiple representations of the same data. If you create and integer
138 and use it as a string, it will be cached in its string form. Sereal
139 attempts to detect the most compact of these representations for
140 encoding, but can not always succeed. For example, if a data structure
141 was previously also traversed by certain other serialization modules
142 (such as Storable), then the scalars in the structure may have been
143 irrevocably upgraded to a more complex (and bigger) type. This is only
144 an issue in crude benchmarks. So if you plan to benchmark
145 serialization, take care not to re-use the test data structure between
146 serializers for results that do not depend on the order of operations.
147
149 For reporting bugs, please use the github bug tracker at
150 <http://github.com/Sereal/Sereal/issues>.
151 For support and discussion of Sereal, there are two Google Groups:
153 Announcements around Sereal (extremely low volume):
155 <https://groups.google.com/forum/?fromgroups#!forum/sereal-announce>
156 Sereal development list:
158 <https://groups.google.com/forum/?fromgroups#!forum/sereal-dev>
159
161 Yves Orton <demerphq@gmail.com>
162 Damian Gryski
164 Steffen Mueller <smueller@cpan.org>
166 Rafaël Garcia-Suarez
168 Ævar Arnfjörð Bjarmason <avar@cpan.org>
170 Tim Bunce
172 Daniel Dragan <bulkdd@cpan.org> (Windows support and bugfixes)
174 Zefram
176 Some inspiration and code was taken from Marc Lehmann's excellent
178 JSON::XS module due to obvious overlap in problem domain.
179
181 This module was originally developed for Booking.com. With approval
182 from Booking.com, this module was generalized and published on CPAN,
183 for which the authors would like to express their gratitude.
184
186 Copyright (C) 2012, 2013, 2014 by Steffen Mueller Copyright (C) 2012,
187 2013, 2014 by Yves Orton
188 The license for the code in this distribution is the following, with
190 the exceptions listed below:
191 This library is free software; you can redistribute it and/or modify it
193 under the same terms as Perl itself.
194 Except portions taken from Marc Lehmann's code for the JSON::XS module,
196 which is licensed under the same terms as this module. (Many thanks to
197 Marc for inspiration, and code.)
198 Also except the code for Snappy compression library, whose license is
200 reproduced below and which, to the best of our knowledge, is compatible
201 with this module's license. The license for the enclosed Snappy code
202 is:
203 Copyright 2011, Google Inc.
205 All rights reserved.
206 Redistribution and use in source and binary forms, with or without
208 modification, are permitted provided that the following conditions are
209 met:
210 * Redistributions of source code must retain the above copyright
212 notice, this list of conditions and the following disclaimer.
213 * Redistributions in binary form must reproduce the above
214 copyright notice, this list of conditions and the following disclaimer
215 in the documentation and/or other materials provided with the
216 distribution.
217 * Neither the name of Google Inc. nor the names of its
218 contributors may be used to endorse or promote products derived from
219 this software without specific prior written permission.
220 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
222 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
223 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
224 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
225 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
226 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
227 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
228 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
229 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
230 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
231 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
232perl v5.32.0 2020-08-04 Sereal::Performance(3)