os/doc
3
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Add sendfile entry from glebius

Browse source
This commit is contained in:
Benjamin Kaduk 2016-01-18 20:13:00 +00:00
parent 76c43ff488
commit e396430b0b
Notes: svn2git 2020-12-08 03:00:23 +00:00 
svn path=/head/; revision=48055
1 changed files with 110 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

110
en_US.ISO8859-1/htdocs/news/status/report-2015-10-2015-12.xml
View file

@ -3631,4 +3631,114 @@
</task>
</help>
</project>
<project cat='kern'>
<title>Sendfile(2) Improvements</title>
<contact>
<person>
<name>
<given>Gleb</given>
<common>Smirnoff</common>
</name>
<email>glebius@FreeBSD.org</email>
</person>
</contact>
<links>
<url href="https://svnweb.FreeBSD.org/base?view=revision&amp;revision=293439">Commit to Head</url>
<url href="http://www.slideshare.net/facepalmtarbz2/new-sendfile-in-english">Slides</url>
<url href="https://events.yandex.ru/lib/talks/2682/">Presentation (in Russian)</url>
</links>
<body>
<p>The <tt>sendfile(2)</tt> system call was introduced in
1998 as an alternative to a traditional
<tt>read(2)</tt>/<tt>write(2)</tt> loop, speeding up server
performance by a factor of ten at the time. Since it was adopted
by all major operating systems, it is now used by any serious web
server software. Wherever there is high traffic, there is
<tt>sendfile(2)</tt> under the hood.</p>
<p>Now, with &os; 11, we are making the next revolutinary step
in serving traffic. <tt>sendfile(2)</tt> no longer blocks waiting
on disk I/O. Instead, it immediately returns control to the
application, performing the necessary I/O in the background. The
original <tt>sendfile(2)</tt> waited for the disk read operation
to complete and then put the data that was read into the socket,
then returned to userspace. If a web server serves thousands of
clients, with thousands of requests, in order to avoid stalls it
needed to spawn extra contexts from which to run
<tt>sendfile(2)</tt>. Alternatively, it could use special tricks
like the <tt>SF_NODISKIO</tt> flag that forces
<tt>sendfile(2)</tt> to serve only content that is cached in
memory. Now, these tricks are in the past, and a web server can
simply use <tt>sendfile(2)</tt> as it would use <tt>write(2)</tt>,
withouth any extra care. The new sendfile cuts out the overhead
of extra contexts, short writes, and extra syscalls to prepopulate
the cache, bringing performance to new level.</p>
<p>The new syscall is built on top of two new features
introduced in the kernel. The first one is an asynchronous VM
pager interface, and the corresponding
<tt>VOP_GETPAGES_ASYNC()</tt> file system method for UFS. The
second one is the concept of &quot;not ready&quot; data in
sockets. When <tt>sendfile(2)</tt> is called, first
VOP_GETPAGES_ASYNC() is called, which dispatches I/O requests for
completion. Then, buffers with pages to be populated are put into
the socket buffer, but flagged as not-yet-ready. Then control
immediately returns to the application. When the I/O is finished,
the buffers are marked as ready, and the socket is activated to
continue transmission.</p>
<p>Additional features of the new sendfile are new flags that
provide an application with extra control over the transmitted
content. Now it is possible to prevent caching of content in
memory, which is useful when it is known that the content is
unlikely to be reused any time soon. In such cases, it is better
to let the associated storage be freed, rather than puting the
data in cache. It is also possible to specify a readahead with
every syscall, if the application can predict client behavior.</p>
<p>The new <tt>sendfile(2)</tt> is a drop-in replacement, API
and ABI compatible with old one. Applications do not even need to
recompile in order to benefit from the new implementation.</p>
<p>This work is a joint effort between two companies: NGINX,
Inc., and Netflix. There were many people involved in the
project. At its initial stage, when no code was yet written, the
idea of such an asynchronous drop-in replacement was discussed
amongst &a.glebius;, &a.scottl;, &a.kib;, &a.adrian;, and Igor
Sysoev. The initial prototype was coded by Gleb under the
supervision of Kostik on the VM parts of patch, and under constant
pressure from Igor, who demanded that nginx be capable of running
with the new <tt>sendfile(2)</tt> with no modifications. The
prototype demonstrated good performance and stability and quickly
went into Netflix production in late 2014. During 2015, the code
matured and continued serving production traffic at Netflix.
&a.scottl;, &a.rrs;, &a.emax;, and &a.gallatin; added their
contributions to the code.</p>
<p>Now we are releasing the code behind our success to the
&os; community, making it available to all &os; users
worldwide!</p>
</body>
<sponsor>
Netflix
</sponsor>
<sponsor>
NGINX, Inc.
</sponsor>
<help>
<task>
<p><tt>SSL_sendfile()</tt> &mdash; an extension to the new
<tt>sendfile(2)</tt> that allows uploading session keys to the
kernel, and then using <tt>sendfile(2)</tt> on an SSL-enabled
socket.</p>
</task>
</help>
</project>
</report>