libpsyc/bench/benchmark.org

#+TITLE: libpsyc Performance Benchmarks

In this document we present the results of performance benchmarks
of libpsyc compared to json-c, libjson-glib, rapidxml and libxml2.

* PSYC, JSON, XML Syntax Benchmarks
First we look at the mere performance of the PSYC syntax
compared to equivalent XML and JSON encodings. We'll
look at actual XMPP messaging later.

** User Profile
In this test we'll compare the efficiency of the three
syntaxes at serializing a typical user data base
storage information. Let's start with XML:

#+INCLUDE: packets/user_profile.xml src xml

In JSON this could look like this:

#+INCLUDE: packets/user_profile.json src js

Here's a way to model this in PSYC (verbose mode):

#+INCLUDE: packets/user_profile.psyc src psyc

** A message with JSON-unfriendly characters
This message contains some characters which are
impractical to encode in JSON. We should probably
put a lot more inside to actually see an impact
on performance.   *TODO*

#+INCLUDE: packets/json-unfriendly.xml src xml
#+INCLUDE: packets/json-unfriendly.json src js
#+INCLUDE: packets/json-unfriendly.psyc src psyc

** A message with XML-unfriendly characters
Same test with characters which aren't practical
in the XML syntax, yet we should put more of
them inside.  *TODO*

#+INCLUDE: packets/xml-unfriendly.xml src xml
#+INCLUDE: packets/xml-unfriendly.json src js
#+INCLUDE: packets/xml-unfriendly.psyc src psyc

** A message with PSYC-unfriendly strings
PSYC prefixes data with length as soon as it
exceeds certain sizes or contains certain strings.
In the case of short messages this is less
efficient than scanning the values without lengths.
Also, lengths are harder to edit by hand.

#+INCLUDE: packets/psyc-unfriendly.xml src xml
#+INCLUDE: packets/psyc-unfriendly.json src js
#+INCLUDE: packets/psyc-unfriendly.psyc src psyc

** Packets containing binary data
We'll use a generator of random binary data to
see how well the formats behave with different
sizes of data. We'll consider 7000 as a possible
size of an icon, 70000 for an avatar, 700000
for a photograph, 7000000 for a piece of music,
70000000 for a large project and
700000000 for the contents of a CD.

* PSYC vs XMPP Protocol Benchmarks

These tests use typical messages from the XMPP ("stanzas" in
Jabber lingo) and compare them with equivalent JSON encodings
and PSYC formats.

** A presence packet
Since presence packets are by far the dominant messaging content
in the XMPP network, we'll start with one of them.
Here's an example from paragraph 4.4.2 of RFC 6121.

#+INCLUDE: packets/presence.xml src xml

And here's the same information in a JSON rendition:

#+INCLUDE: packets/presence.json src js

Here's the equivalent PSYC packet in verbose mode
(since it is a multicast, the single recipients do not
need to be mentioned):

#+INCLUDE: packets/presence.psyc src psyc

And this is the same message in PSYC's compact form, but since compact mode
hasn't been implemented nor deployed yet, you should only consider this
for future projects:

#+INCLUDE: packets/presence-c.psyc src psyc

** An average chat message

#+INCLUDE: packets/chat_msg.xml src xml
#+INCLUDE: packets/chat_msg.json src js
#+INCLUDE: packets/chat_msg.psyc src psyc

Why PSYC doesn't have an id? Because packet counting from contexts
and circuits is automatic: The packet already has a number just by
being there.

Also, PSYC by default doesn't mention a "resource" in XMPP terms,
instead it allows for more addressing schemes than just PSYC.

** A new status updated activity
Example taken from http://onesocialweb.org/spec/1.0/osw-activities.html
You could call this XML namespace hell..  :-)

#+INCLUDE: packets/activity.xml src xml

http://activitystrea.ms/head/json-activity.html proposes a JSON encoding
of this. We'll have to add a routing header to it.

#+INCLUDE: packets/activity.json src js

http://about.psyc.eu/Activity suggests a PSYC mapping for activity
streams. Should a "status post" be considered equivalent to a presence
description announcement or just a message in the "microblogging" channel?
We'll use the latter here:

#+INCLUDE: packets/activity.psyc src psyc

It's nice about XML namespaces how they can by definition never collide,
but this degree of engineering perfection causes us a lot of overhead.
The PSYC approach is to just extend the name of the method - as long as
people use differing method names, protocol extensions can exist next
to each other happily. Method name unicity cannot mathematically be ensured,
but it's enough to append your company name to make it unlikely for anyone
else on earth to have the same name. How this kind of safety is delivered
when using the JSON syntax of ActivityStreams is unclear. Apparently it was
no longer an important design criterion.

* Results

Parsing time of 1 000 000 packets, in milliseconds.
A simple strlen() scan of the respective message is provided for comparison.
These tests were performed on a 2.53 GHz Intel(R) Core(TM)2 Duo P9500 CPU.

|                 | strlen | libpsyc | json-c | json-glib | libxml sax | libxml | rapidxml |
|-----------------+--------+---------+--------+-----------+------------+--------+----------|
| user profile    |     55 |     608 |   4715 |     16503 |       7350 |  12377 |     2477 |
| psyc-unfriendly |     70 |     286 |   2892 |     12567 |       5538 |   8659 |     1896 |
| json-unfriendly |     49 |     430 |   2328 |     10006 |       5141 |   7875 |     1751 |
| xml-unfriendly  |     37 |     296 |   2156 |      9591 |       5571 |   8769 |     1765 |
|-----------------+--------+---------+--------+-----------+------------+--------+----------|
| /               |      < |         |      < |         > |          < |        |        > |
|                 |    <r> |     <r> |    <r> |       <r> |        <r> |    <r> |      <r> |

Pure syntax comparisons above, protocol performance comparisons below:

|          | strlen | libpsyc | libpsyc compact | json-c | json-glib | libxml sax | libxml | rapidxml |
|----------+--------+---------+-----------------+--------+-----------+------------+--------+----------|
| presence |     30 |     236 | 122             |   2463 |     10016 |       4997 |   7557 |     1719 |
| chat msg |     40 |     295 | 258             |   2147 |      9526 |       5911 |   8999 |     1850 |
| activity |     42 |     353 | 279             |   4666 |     16327 |      13357 |  28858 |     4356 |
|----------+--------+---------+-----------------+--------+-----------+------------+--------+----------|
| /        |      < |         | >               |      < |         > |          < |        |        > |
|          |        |         | <c>             |        |           |            |        |          |

Parsing large amounts of binary data. For JSON & XML base64 encoding was used.
Note that the results below include only the parsing time, base64 decoding was
not performed.

|         |   strlen | libpsyc |    json-c |  json-glib | libxml sax |    libxml | rapidxml |
|---------+----------+---------+-----------+------------+------------+-----------+----------|
|      7K |      978 |      77 |     18609 |      98000 |      11445 |     19299 |     8701 |
|     70K |     9613 |      77 |    187540 |    1003900 |      96209 |    167738 |    74296 |
|    700K |    95888 |      77 |   1883500 |   10616000 |     842025 |   1909428 |   729419 |
|      7M |  1347300 |      78 |  26359000 |  120810000 |   12466610 |  16751363 |  7581169 |
|     70M | 14414000 |      80 | 357010000 | 1241000000 |  169622110 | 296017820 | 75308906 |
|---------+----------+---------+-----------+------------+------------+-----------+----------|
|       / |        < |       > |         < |          > |          < |           |        > |
|     <r> |          |         |           |            |            |           |          |

In each case we compared performance of parsing and re-rendering
these messages, but consider also that the applicative processing
of an XML DOM tree is more complicated than just accessing
certain elements in a JSON data structure or PSYC variable mapping.

* Explanations

As you can tell the PSYC data format outpaces its rivals in all circumstances.
Extremely so when delivering binary data as PSYC simply returns the starting
point and the length of the given buffer while the other parsers have to scan
for the end of the transmission, but also with many simpler operations, when
PSYC quickly figures out where the data starts and ends and passes such
information back to the application while the other formats are forced to
generate a copy of the data in order to process possibly embedded special
character sequences. PSYC essentially operates like a binary data protocol
even though it is actually text-based.

* Criticism

Are we comparing apples and oranges? Yes and no, depends on what you
need. XML is a syntax best suited for complex structured data in
well-defined formats - especially good for text mark-up. JSON is a syntax
intended to hold arbitrarily structured data suitable for immediate
inclusion in Javascript source codes. The PSYC syntax is an evolved
derivate of RFC 822, the syntax used by HTTP and E-Mail. It is currently
limited in the kind and depth of data structures that can be represented
with it, but it is highly efficient in exchange.

In fact we are currently looking into suitable syntax extensions to represent
generic structures and semantic signatures, but for now PSYC only
provides for simple typed values and lists of typed values.

* Ease of Implementation

Another aspect is the availability of these formats for spontaneous
use. You could generate and parse JSON yourself but you have to be
careful about escaping. XML can be rendered manually if you know your
data will not break the syntax, but you shouldn't dare to parse it without
a bullet proof parser. PSYC is easy to render and parse yourself for
simple tasks, as long as the body does not contain "\n|\n" and your
variables do not contain newlines.

* Conclusions

After all it is up to you to find out which format fulfils your
requirements the best. We use PSYC for the majority of messaging where
JSON and XMPP aren't efficient and opaque enough, but we employ XML and
JSON as payloads within PSYC for data that doesn't fit the PSYC model.
For some reason all three formats are being used for messaging, although
only PSYC was actually designed for that purpose.

The Internet has developed two major breeds of protocol formats.
The binary ones are extremely efficient but in most cases you have
to recompile all instances each time you change something
while the plain-text ones are reaching out for achieving perfection
in data representation while leaving the path of efficiency. Some
protocols such as HTTP and SIP are in-between these two schools,
offering both a text-based extensible syntax (it's actually easier to
add a header to HTTP than to come up with a namespace for XMPP...)
and the ability to deliver binary data. But these protocols do not
come with native data structure support. PSYC is a protocol that
combines the compactness and efficiency of binary protocols with the
extensibility of text-based protocols and still provides for enough
data structuring to rarely require the use of other data formats.

* Futures

After a month of development libpsyc is already performing pretty
well, but we presume various optimizations, like rewriting parts
in assembler, are possible.

* Related Work

If this didn't help, you can also look into:

- Adobe AMF
- ASN.1
- BSON
- Cisco Etch
- Efficient XML
- Facebook Thrift
- Google Protocol Buffers

The drawback of these binary formats is, unlike PSYC, JSON and XML
you can't edit them manually and you can't produce valid messages
by replacing variables in a simple text template. You depend on
specialized parsers and renderers to be provided.

There's also

- Bittorrent's bencode

This format is formally text-based, but not easy to read as it doesn't
have any visual separators and isn't easy to edit as everything is
prefixed by lengths even for very short items.

* Further Reading

http://about.psyc.eu/Spec:Syntax provides you with the ABNF grammar
of the PSYC 1.0 syntax. You may also be interested in PSYC's decentralized
state mechanism provided by the +/-/= operators.

See http://about.psyc.eu/XML and http://about.psyc.eu/JSON for more
biased information on the respective formats.

* Appendix
** Tools used

This document and its benchmarks are distributed with libpsyc.
See http://about.psyc.eu/libpsyc on how to obtain it.

The benchmarks can be run with the following command
(xmlbench is needed for the xml tests):

: make bench