Hi Xavier, Any plan to put it in alpha2? It doesn't look like it requires a lot of work if we just want to change the buffer size. nio may be a long shot.
Thanks Xavier Hanin wrote: > > Great, thanks a lot for these data! > > Xavier > > On 6/14/07, Buck, Robert <[EMAIL PROTECTED]> wrote: >> >> What the performance test did is: >> >> 1. increase by powers of 2 a file size variable >> >> 2. for each file size, create/open a new file >> >> 3. write specified number of bytes to file >> >> 4. close file >> >> Tests were performed on two systems: >> >> 1. IBM ThinkPad T42 (slow) laptop, stock hardware, 1.5GB RAM >> >> 2. Dell Dual CPU, 5150, 64-bit Red Hat 4; (2) SAS 15k 146GB disks in >> RAID 1 configuration; 16 GB RAM; this is production grade hardware. >> >> Both sytems produced similar profiles (shape), however, on the >> production hardware with Linux (ext2, I believe) the numbers were >> substantially greater (makes sense when you consider the fast disks I >> have; at 3GB/sec rates, these SAS drives scream). >> >> In both cases buffers were configured for both Old IO and New IO. With >> NIO direct buffers were used. In both cases, I determined the ideal >> buffer sizes based upon experimentation to be around 32kb - 64 kb. Going >> to buffer sizes less than, or greater than, this range reduced overall >> throughput. >> >> -Bob >> >> > -----Original Message----- >> > From: Xavier Hanin [mailto:[EMAIL PROTECTED] >> > Sent: Thursday, June 14, 2007 8:38 AM >> > To: [email protected] >> > Subject: Re: ivy:retrieve performance >> > >> > On 6/14/07, Buck, Robert <[EMAIL PROTECTED]> wrote: >> > > >> > > Folks, >> > > >> > > We benchmarked a number of JDK IO API's for an internal project. To >> > > neutralize any questions regarding NIO vs Old IO, please >> > take a look >> > > at the attached diagram. These rates will be largely >> > identical on both >> > > Linux and Windows. Blue line NIO, red line old io. >> > >> > >> > I've uploaded the image here: >> > http://www.flickr.com/photos/[EMAIL PROTECTED]/548163404/ >> > >> > Could you give a little bit more details about how you did >> > the tests: jvm used, size of buffer used for old IO, ... >> > >> > According to your tests it seems that NIO should be preferred >> > in any case, it wasn't what some users in the javalobby >> > thread seemed to say. So I wonder what makes the difference. >> > >> > Xavier >> > >> > If you do not get the attached JPEG file, let me know. I can send it >> > > directly to you if you so request. >> > > >> > > Cheers, >> > > >> > > Bob >> > > >> > > > -----Original Message----- >> > > > From: Xavier Hanin [mailto:[EMAIL PROTECTED] >> > > > Sent: Thursday, June 14, 2007 6:22 AM >> > > > To: [email protected] >> > > > Subject: Re: ivy:retrieve performance >> > > > >> > > > On 6/14/07, Gilles Scokart <[EMAIL PROTECTED]> wrote: >> > > > > >> > > > > > 2. The buffer in FileUtils.java is too small. It's set at >> > > > 8192. It >> > > > > > seems to >> > > > > > > be much better for me to set it much larger. This is due to >> > > > > > > the fact >> > > > > > that >> > > > > > > it >> > > > > > > needs to read and write simultaneously. The bigger the >> > > > buffer is, >> > > > > > > the smaller number of time, HD header has to move. For >> > > > me, 65536 >> > > > > > > seems to perform much better but I haven't tried >> > other numbers. >> > > > > > >> > > > > > >> > > > > > I'd like to get more feedback on this. One use case is >> > > > not the other. >> > > > > This >> > > > > > size has been borrowed from Ant copy mechanism. Maybe >> > > > what we could >> > > > > > do >> > > > > is >> > > > > > make this configurable, so that one could adapt to its >> > > > needs. Or try >> > > > > > to guess a good size depending on the size (when it's >> > possible >> > > > > > to get an >> > > > > idea >> > > > > > of the size before copying). >> > > > > > >> > > > > > Xavier >> > > > > > >> > > > > >> > > > > >> > > > > Couldn't we use the nio for that? (See >> > > > > http://www.javalobby.org/java/forums/t17036.html) >> > > > >> > > > >> > > > According to comments 10 and 11 NIO have bad performance >> > for large >> > > > files on linux, and input stream with byte buffer is >> > pretty close to >> > > > NIO for small files (see comment 13 conclusion). So I'm not sure >> > > > switching to NIO would indeed help a lot. According to >> > the tests in >> > > > the thread you pointed using a 64kB buffer seems to be a >> > good choice >> > > > (which confirms testn tests), at least for large files. OTOH the >> > > > last conclusion (comment 17) is different. >> > > > So I don't really know what to think about that. We >> > should make some >> > > > tests on several platforms and jvms to draw conclusion >> > ourself, but >> > > > it takes time. >> > > > >> > > > >> > > > Xavier >> > > > >> > > > Gilles >> > > > > >> > > > > >> > > > > >> > > > > >> > > > >> > > > >> > > > -- >> > > > Xavier Hanin - Independent Java Consultant Manage your >> > dependencies >> > > > with Ivy! >> > > > http://incubator.apache.org/ivy/ >> > > > >> > > >> > >> > >> > >> > -- >> > Xavier Hanin - Independent Java Consultant Manage your >> > dependencies with Ivy! >> > http://incubator.apache.org/ivy/ >> > >> > > > > -- > Xavier Hanin - Independent Java Consultant > Manage your dependencies with Ivy! > http://incubator.apache.org/ivy/ > > -- View this message in context: http://www.nabble.com/ivy%3Aretrieve-performance-tf3907253.html#a11230918 Sent from the ivy-user mailing list archive at Nabble.com.
