Author: pgj
Date: Thu Jan 9 19:08:06 2014
New Revision: 43458
URL: http://svnweb.freebsd.org/changeset/doc/43458
Log:
- Add a 2013Q4 report entry on ZFS/NFS improvements
Submitted by: mav
Modified:
head/en_US.ISO8859-1/htdocs/news/status/report-2013-10-2013-12.xml
Modified: head/en_US.ISO8859-1/htdocs/news/status/report-2013-10-2013-12.xml
==============================================================================
--- head/en_US.ISO8859-1/htdocs/news/status/report-2013-10-2013-12.xml Thu Jan 9 16:51:55 2014 (r43457)
+++ head/en_US.ISO8859-1/htdocs/news/status/report-2013-10-2013-12.xml Thu Jan 9 19:08:06 2014 (r43458)
@@ -19,7 +19,7 @@
<!-- XXX: Keep updating the number of entries -->
<p>Thanks to all the reporters for the excellent work! This report
- contains 9 entries and we hope you enjoy reading it.</p>
+ contains 10 entries and we hope you enjoy reading it.</p>
<p>The deadline for submissions covering between January and
March 2014 is April 7th, 2014.</p>
@@ -38,6 +38,12 @@
</category>
<category>
+ <name>kern</name>
+
+ <description>Kernel</description>
+ </category>
+
+ <category>
<name>arch</name>
<description>Architectures</description>
@@ -498,4 +504,59 @@
reports!</task>
</help>
</project>
+
+ <project cat='kern'>
+ <title>UMA/ZFS and RPC/NFS Performance Improvements</title>
+
+ <contact>
+ <person>
+ <name>
+ <given>Alexander</given>
+ <common>Motin</common>
+ </name>
+ <email>mav at FreeBSD.org</email>
+ </person>
+ </contact>
+
+ <links>
+ <url href="http://docs.freebsd.org/cgi/mid.cgi?52894C92.60905">Discussion of the ZFS/UMA changes</url>
+ </links>
+
+ <body>
+ <p>The performance of ZFS and NFS was suboptimal in &os;, so we
+ have recently investigated some possible improvement paths. The
+ <tt>uma(9)</tt> memory allocator caching code was improved to
+ adapt better to system memory constraints. Combined with other
+ virtual memory subsystem improvements done in the previous
+ years, it should be safe to actively use <tt>uma(9)</tt> caches
+ now. It may be enabled for ZFS to use them for ZIO/ARC via the
+ <tt>vfs.zfs.zio.use_uma</tt> <tt>loader(8)</tt> tunable, which
+ is now the default for amd64, where it is recommended. Use of
+ <tt>uma(9)</tt> caches for LZ4 compression buffers is
+ unconditionally enabled on all architectures as it is has no
+ serious drawbacks. On systems with many CPUs, these changes
+ doubled the performance in the benchmarks.</p>
+
+ <p>Several areas of the NFS server stack (RPC, FHA, DRC) got
+ number of fixes and performance optimizations that significantly
+ improves the performance and reduces the CPU usage in number of
+ tests. Together with the ZFS memory allocator changes mentioned
+ above, it was possible to reach 200K NFS block read IOPS and 55K
+ SPEC NFS IOPS.</p>
+
+ <p>The code was committed to <tt>head</tt>. The <tt>uma(9)</tt>
+ ZFS commits have been already merged to <tt>stable/10</tt>, and
+ the remainder will be done soon as well.</p>
+
+ <p>The project is sponsored by iXsystems, Inc.</p>
+ </body>
+
+ <help>
+ <task>The SPEC NFS test hits lock congestion on several global
+ locks in the file system layer when a quite intensive
+ <tt>READDIRPLUS</tt> NFS request is received. Fixing this
+ problem could improve the performance on large systems even
+ further.</task>
+ </help>
+ </project>
</report>