proposal to optimise the performance of the Jar utility
mike.skells at talk21.com
mike.skells at talk21.com
Thu Apr 14 08:54:22 UTC 2011
Hi Sherman,
I agree that the update case is different, but always thought that your new zip
file system would be a better way to do this rather than unpacking and repacking
a jar, which is what happens not isnt it? Forgive me if I am wrong but I never
looked at the code in detail for update
I will check up on the timing that I get, but it will be after the weekend. It
may be that they were skewed by the overhead of the profiler etc. I was running
with full instrumentation for those timings
Just to be clear I was saying that the CPU time taken was by the hashcode was
greater than the CPU time taken in deflate. I am attempting to eliminate any
real IO time for the timings
I have also started on pipelining the 'expand' and 'create' code sections so
that the expand time can be eliminated, if there is no IO contention, and
another core available.
In this case there is no list stored, just a queue of files to be added
For the cf0 case the quick win is not scanning the file twice (once for Adler32
and once for content), which in my tests takes a couple of seconds off the
create time
regards
Mike
________________________________
From: Xueming Shen <xueming.shen at oracle.com>
To: mike.skells at talk21.com
Cc: core-libs-dev Libs <core-libs-dev at openjdk.java.net>
Sent: Thursday, 14 April, 2011 7:12:26
Subject: Re: proposal to optimise the performance of the Jar utility
On 4/13/2011 4:55 PM, mike.skells at talk21.com wrote:
Hi Mike,
>apart from compression 0 it doesn't make a lot of difference as far as
>I see it. As I posted earlier, converting the file name to lower case
>seems (according the the NB profiler) to take more time than the
>compression
File.hashCode() is slow on Windows (yes, we have to go down to
String.toLowerCase() to be "correct" and yes, its slow), we have
the comment in Win32FileSystem.hashCode() just for this:-)
public int hashCode(File f) {
/* Could make this more efficient: String.hashCodeIgnoreCase */
return f.getPath().toLowerCase(Locale.ENGLISH).hashCode() ^ 1234321;
}
I tried to use ArrayList for the "entries" (instead of LinkedHashSet), as
showed below
(yes, it's total fine to use List for "creating", need a little more work
for "updating", but
let's put "updating" aside for now)
// All files need to be added/updated.
//Set<File> entries = new LinkedHashSet<File>();
List<File> entries = new ArrayList<File>();
and run the cf0 on rtjar, which has the rt.jar extracted (time measured by
System.currentTimeMillis()
[List]
$ ../jdk1.7.0/bin/java Main cf0 foo.jar rtjar
expand=406, create=10703
expand=406, create=10797
expand=407, create=11046
[Set]
$ ../jdk1.7.0/bin/java Main cf0 foo.jar rtjar
expand=469, create=10703
expand=469, create=10688
expand=469, create=10656
The "List" version is indeed about 15% faster than the "Set" version, in
"expanding", but
"expanding" is only about 5% of the over all "creating" time (I've not tried
to measure
yet, but I will not be surprised to find that we probably spend more time on
i/o than deflating)
So it is nice to have, but does not appear to bring us some thing
"significant".
-Sherman
>
>Bearing in mind that I have tweaked my system so that I am accessing
>from cache only, so the elapsed time is pretty much the single core
>time, here are the results
>
>
>the result I have re-run today (hence the delay)
>in csv format
>---
>,cf,cf0,cf1,cf2,cf3,cf4,cf5,cf6,cf7,cf8,cf9
>1.6.0_24 (1.6 vm),10255,9751,,,,,,,,,
>java 1.7 release candinate,10498,9663,,,,,,,,,
>orig ,10481,9707,,,,,,,,,
>buffer CRC32
>data,,7398,9490,9641,9565,10038,10142,10366,10310,10536,10440
>---
>this run shows a 25% improvement in cf0 times (9703 ms vs 7398 wthen
>the tweaks are applied)
>cf1 (compressed level 1) takes 9490 and cf9 takes 10440, so not a huge
>difference
>
>
>
>
>Regards
>
>
>Mike
>
>
>
>
>
>
>
________________________________
From: Mike Duigou <mike.duigou at oracle.com>
>To: mike.skells at talk21.com
>Cc: Xueming Shen <xueming.shen at oracle.com>; core-libs-dev Libs
><core-libs-dev at openjdk.java.net>
>Sent: Wednesday, 13 April, 2011 23:00:26
>Subject: Re: proposal to optimise the performance of the Jar
>utility
>
>Mike, can you share the results of performance testing at various
>compression levels? Is there much difference between the levels or
>an apparent "sweet spot"?
>
>For low hanging fruit for jdk 7 it might be worth considering
>raising the default compression level from 5 to 6 (the zlib
>default). Raising the level from 5 to 6 entails (by today's
>standards) very modest increases in the amount of memory and effort
>used (16Kb additional buffer space for compression). In general zlib
>reflects size choices that are almost 20 years old and it may be of
>no measurable benefit to be using the lower compression levels.
>
>Mike (also)
>
>
>On Apr 12 2011, at 17:48 , mike.skells at talk21.com wrote:
>
>> Hi Sherman,
>> I have had a quick look at the current code to see what 'low
>>hanging fruit'
>>
>> there is. I appreciate that parallelizing the command in its
>>entirity may not be
>>
>> feasible for the first release
>>
>> The tests that I have run are jarring the content of the 1.7 rt.jar
>>with varying
>>
>> compression levels. Each jar is run as an child process 6 times and
>>the average
>>
>> of the last 5 is taken. Tests are pretty much CPU bound on a single
>>core
>>
>> 1. The performance of the cf0 (create file with no compression)
>>path can be
>>
>> improved for the general case if the file is buffered.
>> In my test env (windows 7 64bit) this equates to a 17% time
>>performance
>>
>> improvement in my tests. In the existing code the file is read
>>twice, once to
>>
>> calc the CRC and once to write the file to the Jar file. This
>>change would
>>
>> buffer a single small file at a time (size < 100K)
>>
>> 2. It is also a trivial fix to allow different compression levels,
>>rather than
>>
>> stored and default
>>
>> After that it is harder to gain performance improvements without
>>structural
>>
>> change or more discussion
>>
>> 3. The largest saving after that is in the management of the
>>'entries' Set, as
>>
>> the hashcode of the File is expensive (this may not apply to other
>>filesystems).
>>
>> the management of this map seems to account for more cpu than
>>Deflator!
>> I cannot see the reason for this data being collected. I am
>>probably missing the
>>
>> obvious ...
>>
>> 4. After that there is just the parallelisation of the jar utility
>>and the
>>
>> underlying stream
>>
>> What is the best way to proceed
>>
>> regards
>>
>> Mike
>>
>>
>>
>> ________________________________
>> From: Xueming Shen <xueming.shen at oracle.com>
>> To: core-libs-dev at openjdk.java.net
>> Sent: Wednesday, 6 April, 2011 19:04:25
>> Subject: Re: proposal to optimise the performance of the Jar
>>utility
>>
>> Hi Mike,
>>
>> We are in the final stage of the JDK7 development, work like this
>>is
>>
>> unlikely to get in the
>> last minute. I have filed a CR/RFE to trace this issue, we can use
>>this
>>
>> CR as the start
>> point for the discussion and target for some jdk7 update releases
>>or JDK8.
>>
>> 7034403: proposal to optimise the performance of the Jar utility
>>
>> Thanks,
>> Sherman
>>
>>
>> On 04/05/2011 04:42 PM, mike.skells at talk21.com wrote:
>>> Hi,
>>> Not sure if this is too late for Java 7 but I have made some
>>>optimisations for
>>>
>>> a
>>> client to improve the performance of the jar utility in their
>>>environment, and
>>> would like to promite then into the main code base
>>>
>>> The optimisations that I have performed are
>>>
>>> 1. Allowing the Jar utility to have other compression levels
>>>(currently it
>>> allows default (5) only)
>>> 2. Multi-threading, and pipelining the the file information and
>>>access
>>> 3. multi-threading the compression and file writing
>>>
>>> A little background
>>> A part of the development process of where I work they regularly
>>>Jar the
>>>
>>> content
>>> of the working projects as part of the distribution to remote
>>>systems. This is
>>>
>>> a
>>> large and complex code base of 6 million LOC and growing. The Jar
>>>file ends up
>>> compressed to approx 100Mb, Uncompressed the jar size is approx
>>>245mb, about
>>>
>>> 4-5
>>> times the size of rt.jar.
>>>
>>> I was looking at ways to improve the performance as this activity
>>>occurs
>>>
>>> several
>>> times a day for dozens of developers
>>>
>>> In essence when compressing a new jar file the jar utility is
>>>single threaded
>>> and staged. Forgive me if this is an oversimplification
>>>
>>> first it works out all of the files that are specified, buffering
>>>the file
>>> names, (IO bound)
>>> then it iterates through the files, and for each file, it load the
>>>file
>>> information, and then the file content sending it to a
>>>JarOutputStream, (CPU
>>> bound or IO bound depending on the IO speed)
>>>
>>> The JarOutputStream has a compression of 0 (just store) or 5 (the
>>>default),
>>>
>> and
>>> the jar writing is single threaded by the design of the
>>>JarOutputStream
>>>
>>> The process of creation of a Jar took about 20 seconds in windows
>>>with the
>>>
>> help
>>> of an SSD, and considerable longer without one, and was CPU bound
>>>to one CPU
>>> core
>>>
>>> ----
>>> The changes that I made were
>>> 1. Allow deferent compression levels (for us a compression level of
>>>1
>>>
>> increases
>>> the file size of the Jar to 110 Mb but reduces the CPU load in
>>>compression to
>>> approx 30% of what it was (rough estimate)
>>> 2. pipelining the file access
>>> 2.1 one thread is started for each file root (-C on the Jar
>>>command line),
>>> which scans for files and places the file information into a
>>>blocking
>>>
>>> queue(Q1),
>>> which I set to abretrary size of 200 items
>>> 2.2 one thread pool of 10 threads reads the file information
>>>from the queue
>>> (Q1) and buffers the file content to a specified size (again I
>>>specified an
>>> arbetrary size limit of 25K for a file, and places the buffered
>>>content into a
>>> queue(q2) (again arbetrary size of 10 items
>>> 2.3 one thread takes the filecontent from Q2 and compresses it
>>>or checksums
>>> it and adds it the the JarOutputStream. This process is single
>>>threaded due
>>>
>> to
>>> the design of the JarOutputStream
>>>
>>> some other minor performance gain occurred by increasing the buffer
>>>on the
>>> output stream to reduce the IO load
>>>
>>> The end result is that the process takes about approx 5 seconds in
>>>the same
>>> configuration
>>>
>>> The above is in use in production configuration for a few months
>>>now
>>>
>>> As a home project I have completed some enhancements to the
>>>JarOutputStream,
>>>
>>> and
>>> produced a JarWriter that allows multiple threads to work
>>>concurrently
>>>
>>> deflating
>>> or calculating checksums, which seems to test OK for the test cases
>>>that Ihave
>>> generated,and successfully loads my quad core home dev machine on
>>>all cores.
>>> Each thread allocates a buffer, and the thread compresses a files
>>>into the
>>> buffer, only blocking other threads whenthe buffer is written to
>>>the output
>>> (which is after the compression is complete, unless the file is too
>>>large to
>>> compress
>>>
>>> This JarWriter is not API compatable with the JarOutputStream, it
>>>is not a
>>> stream. It allows the programmer to write a record based of the
>>>file
>>>
>>> information
>>> and an input stream, and is threadsafe. It is not a drop in
>>>replacement for
>>> JarOutputStream
>>> I am not an expert in the ZIp file format, but much of the code
>>>from
>>> ZipOutputStream is unchanged, just restructured
>>> ---
>>> I did think that there is some scope for improvement, that I have
>>>not looked
>>>
>> at
>>> a. thresholding of file size for compression (very small files dont
>>>compress
>>> well
>>> b. some file types dont compress well (e.g. png, jpeg) as they have
>>>been
>>> compressed already)
>>> c. using NIO to parallelise the loading of the file information or
>>>content
>>> d. some pre-charging of the deflator dictionary (e.g. a class file
>>>contains
>>>
>> the
>>> strings of the class name and packages), but this would make the
>>>format
>>> incompatable with zip, and require changes to the JVM to be useful,
>>>and is a
>>> long way from my comform zone, or skill set. This would reduce the
>>>file size
>>>
>>> --
>>> What is the view of the readers. Is this something, or at least
>>>some parts of
>>> this that could be incorperated into Java 7 or is this too late on
>>>the dev
>>>
>>> cycle
>>>
>>> regards
>>>
>>> Mike
>
>
More information about the core-libs-dev
mailing list