NAME AnyEvent::Net::Curl::Queued - Any::Moose wrapper for queued downloads via Net::Curl & AnyEvent VERSION version 0.041 SYNOPSIS #!/usr/bin/env perl package CrawlApache; use feature qw(say); use strict; use utf8; use warnings qw(all); use HTML::LinkExtor; use Any::Moose; extends 'AnyEvent::Net::Curl::Queued::Easy'; after finish => sub { my ($self, $result) = @_; say $result . "\t" . $self->final_url; if ( not $self->has_error and $self->getinfo('content_type') =~ m{^text/html} ) { my @links; HTML::LinkExtor->new(sub { my ($tag, %links) = @_; push @links, grep { $_->scheme eq 'http' and $_->host eq 'localhost' } values %links; }, $self->final_url)->parse(${$self->data}); for my $link (@links) { $self->queue->prepend(sub { CrawlApache->new($link); }); } } }; no Any::Moose; __PACKAGE__->meta->make_immutable; 1; package main; use strict; use utf8; use warnings qw(all); use AnyEvent::Net::Curl::Queued; my $q = AnyEvent::Net::Curl::Queued->new; $q->append(sub { CrawlApache->new('http://localhost/manual/') }); $q->wait; DESCRIPTION AnyEvent::Net::Curl::Queued (a.k.a. YADA, *Yet Another Download Accelerator*) is an efficient and flexible batch downloader with a straight-forward interface capable of: * create a queue; * append/prepend URLs; * wait for downloads to end (retry on errors). Download init/finish/error handling is defined through Moose's method modifiers. MOTIVATION I am very unhappy with the performance of LWP. It's almost perfect for properly handling HTTP headers, cookies & stuff, but it comes at the cost of *speed*. While this doesn't matter when you make single downloads, batch downloading becomes a real pain. When I download large batch of documents, I don't care about cookies or headers, only content and proper redirection matters. And, as it is clearly an I/O bottleneck operation, I want to make as many parallel requests as possible. So, this is what CPAN offers to fulfill my needs: * Net::Curl: Perl interface to the all-mighty libcurl , is well-documented (opposite to WWW::Curl); * AnyEvent: the DBI of event loops. Net::Curl also provides a nice and well-documented example of AnyEvent usage (03-multi-event.pl); * MooseX::NonMoose: Net::Curl uses a Pure-Perl object implementation, which is lightweight, but a bit messy for my Moose-based projects. MooseX::NonMoose patches this gap. AnyEvent::Net::Curl::Queued is a glue module to wrap it all together. It offers no callbacks and (almost) no default handlers. It's up to you to extend the base class AnyEvent::Net::Curl::Queued::Easy so it will actually download something and store it somewhere. ALTERNATIVES As there's more than one way to do it, I'll list the alternatives which can be used to implement batch downloads: * WWW::Mechanize: no (builtin) parallelism, no (builtin) queueing. Slow, but very powerful for site traversal; * LWP::UserAgent: no parallelism, no queueing. WWW::Mechanize is built on top of LWP, by the way; * LWP::Protocol::Net::Curl: *drop-in* replacement for LWP::UserAgent, WWW::Mechanize and their derivatives to use Net::Curl as a backend; * LWP::Curl: LWP::UserAgent-alike interface for WWW::Curl. Not a *drop-in*, no parallelism, no queueing. Fast and simple to use; * HTTP::Tiny: no parallelism, no queueing. Fast and part of CORE since Perl v5.13.9; * HTTP::Lite: no parallelism, no queueing. Also fast; * Furl: no parallelism, no queueing. Very fast, despite being pure-Perl; * Mojo::UserAgent: capable of non-blocking parallel requests, no queueing; * AnyEvent::Curl::Multi: queued parallel downloads via WWW::Curl. Queues are non-lazy, thus large ones can use many RAM; * Parallel::Downloader: queued parallel downloads via AnyEvent::HTTP. Very fast and is pure-Perl (compiling event driver is optional). You only access results when the whole batch is done; so huge batches will require lots of RAM to store contents. BENCHMARK (see also: CPAN modules for making HTTP requests ) Obviously, every download agent is (or, ideally, should be) *I/O bound*. However, it is not uncommon for large concurrent batch downloads to hog the processor cycles before consuming the full network bandwidth. The proposed benchmark measures the request rate of several concurrent download agents, trying hard to make all of them *CPU bound* (by removing the I/O constraint). On practice, this benchmark results mean that download agents with lower request rate are less appropriate for parallelized batch downloads. On the other hand, download agents with higher request rate are more likely to reach the full capacity of a network link while still leaving spare resources for data parsing/filtering. The script eg/benchmark.pl compares AnyEvent::Net::Curl::Queued (A.K.A. YADA) against several other download agents. Only AnyEvent::Net::Curl::Queued itself, AnyEvent::Curl::Multi, Parallel::Downloader, Mojo::UserAgent and lftp support concurrent downloads natively; thus, Parallel::ForkManager is used to reproduce the same behaviour for the remaining agents, while taskset avoids the skew on multiprocessor systems. The download target is a copy of the Apache documentation on a local Apache server. The test platform configuration: * Intel® Core™ i7-2600 CPU @ 3.40GHz with 8 GB RAM; * Ubuntu 11.10 (64-bit); * Perl v5.16.2 (installed via perlbrew); * libcurl/7.28.0 (without AsynchDNS, which slows down curl_easy_init() ). The script eg/benchmark.pl uses Benchmark::Forking and Class::Load to keep UA modules isolated and loaded only once. $ taskset 1 perl benchmark.pl --count 100 --parallel 8 --repeat 10 Request rate WWW::M LWP::UA L::P::N::C Mojo::UA HTTP::L HTTP::T lftp P::D AE::C::M YADA Furl curl wget LWP::C WWW::Mechanize v1.72 534/s -- -32% -61% -63% -80% -82% -83% -84% -85% -86% -94% -95% -97% -97% LWP::UserAgent v6.04 782/s 46% -- -42% -46% -71% -73% -75% -76% -77% -79% -92% -93% -95% -95% LWP::Protocol::Net::Curl v0.011 1360/s 154% 74% -- -6% -50% -53% -57% -59% -61% -64% -86% -88% -91% -91% Mojo::UserAgent v3.82 1450/s 171% 85% 7% -- -46% -50% -54% -56% -58% -62% -85% -87% -91% -91% HTTP::Lite v2.4 2700/s 405% 245% 98% 86% -- -7% -14% -18% -22% -29% -71% -76% -82% -83% HTTP::Tiny v0.025 2910/s 445% 272% 114% 101% 8% -- -7% -11% -16% -23% -69% -74% -81% -81% lftp v4.3.1 3140/s 488% 302% 131% 117% 17% 8% -- -4% -9% -17% -67% -72% -80% -80% Parallel::Downloader v0.121560 3280/s 514% 319% 141% 127% 22% 13% 4% -- -5% -13% -65% -70% -79% -79% AnyEvent::Curl::Multi v1.1 3460/s 548% 342% 155% 139% 28% 19% 10% 5% -- -9% -63% -69% -77% -78% YADA v0.038 3790/s 610% 385% 179% 162% 41% 30% 21% 16% 10% -- -60% -66% -75% -76% Furl v2.01 9420/s 1663% 1104% 593% 550% 249% 223% 200% 187% 172% 148% -- -15% -39% -40% curl v7.28.0 11100/s 1977% 1318% 716% 666% 311% 281% 253% 238% 221% 193% 18% -- -28% -29% wget v1.12 15400/s 2777% 1864% 1031% 961% 470% 428% 389% 368% 344% 305% 63% 39% -- -1% LWP::Curl v0.12 15600/s 2818% 1892% 1047% 976% 478% 435% 396% 375% 350% 311% 65% 40% 1% -- (output formatted to show module versions at row labels and keep column labels abbreviated) ATTRIBUTES allow_dups Allow duplicate requests (default: false). By default, requests to the same URL (more precisely, requests with the same signature are issued only once. To seed POST parameters, you must extend the AnyEvent::Net::Curl::Queued::Easy class. Setting "allow_dups" to true value disables request checks. common_opts "opts" in AnyEvent::Net::Curl::Queued::Easy attribute common to all workers initialized under the same queue. You may define "User-Agent" string here. http_response Encapsulate the response with HTTP::Response (only when the scheme is HTTP/HTTPS); a global version of "http_response" in AnyEvent::Net::Curl::Queued::Easy. Default: disabled. completed Count completed requests. cv AnyEvent condition variable. Initialized automatically, unless you specify your own. Also reset automatically after "wait", so keep your own reference if you really need it! max Maximum number of parallel connections (default: 4; minimum value: 1). multi Net::Curl::Multi instance. queue "ArrayRef" to the queue. Has the following helper methods: queue_push Append item at the end of the queue. queue_unshift Prepend item at the top of the queue. dequeue Shift item from the top of the queue. count Number of items in queue. share Net::Curl::Share instance. stats AnyEvent::Net::Curl::Queued::Stats instance. timeout Timeout (default: 60 seconds). unique Signature cache. watchdog The last resort against the non-deterministic chaos of evil lurking sockets. METHODS start() Populate empty request slots with workers from the queue. empty() Check if there are active requests or requests in queue. add($worker) Activate a worker. append($worker) Put the worker (instance of AnyEvent::Net::Curl::Queued::Easy) at the end of the queue. For lazy initialization, wrap the worker in a "sub { ... }", the same way you do with the Moose "default => sub { ... }": $queue->append(sub { AnyEvent::Net::Curl::Queued::Easy->new({ initial_url => 'http://.../' }) }); prepend($worker) Put the worker (instance of AnyEvent::Net::Curl::Queued::Easy) at the beginning of the queue. For lazy initialization, wrap the worker in a "sub { ... }", the same way you do with the Moose "default => sub { ... }": $queue->prepend(sub { AnyEvent::Net::Curl::Queued::Easy->new({ initial_url => 'http://.../' }) }); wait() Process queue. CAVEAT * Many sources suggest to compile libcurl with c-ares support. This only improves performance if you are supposed to do many DNS resolutions (e.g. access many hosts). If you are fetching many documents from a single server, "c-ares" initialization will actually slow down the whole process! SEE ALSO * AnyEvent * Any::Moose * Net::Curl * WWW::Curl * AnyEvent::Curl::Multi AUTHOR Stanislaw Pusep COPYRIGHT AND LICENSE This software is copyright (c) 2013 by Stanislaw Pusep. This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.