I added some basic JMH tests to my repo along with a couple alternative appender implementations. I got rid of the unnecessary file region locking in the async file channel one, but it's still coming out quite a bit slower than the RandomAccessFile and Files.newOutputStream() based appenders, though that could be due to the use of Phaser (which I only added to cleanly close the appender synchronously).
On 26 February 2017 at 10:05, Matt Sicker <boa...@gmail.com> wrote: > Perhaps something got optimized by the JVM? I'll add some JMH tests to > this repo to try out various approaches. > > On Sat, Feb 25, 2017 at 21:12, Apache <ralph.go...@dslextreme.com> wrote: > >> I tried using a FileChannel for the FileAppender a week or so ago to see >> if passing the ByteBuffer to the FileChannel would improve performance >> since it doesn’t have to be synchronized. I didn’t see any improvement >> though and I ended up reverting it. But I might have done something wrong. >> >> Ralph >> >> On Feb 25, 2017, at 4:19 PM, Matt Sicker <boa...@gmail.com> wrote: >> >> We already use a bit of NIO (ByteBuffer for layouts and >> appenders/managers, MappedByteBuffer for mmap'd files, FileLock for locking >> files, etc.), and I've been playing around with the NIO API lately. I have >> some sample code here <https://github.com/jvz/nio-logger> to show some >> trivial use case of AsynchronousFileChannel. In Java 7, there is also >> AsynchronousSocketChannel which could theoretically be used instead of >> adding Netty for a faster socket appender. In that regard, I'm curious as >> to how useful it would be to have similar appenders as the OutputStream >> ones, but instead using WritableByteChannel, GatheringByteChannel (possible >> parallelization of file writing?), and the async channels (there's an >> AsynchronousByteChannel class, but I think they screwed this one up as only >> one of the three async channel classes implements it). >> >> Another related issue I've seen is that in a message-oriented appender >> (e.g., the Kafka one), being able to stream directly to a ByteBuffer is not >> the right way to go about encoding log messages into the appender. Instead, >> I was thinking that a pool of reusable ByteBuffers could be used here where >> a ByteBuffer is borrowed on write and returned on completion (via a >> CompletionHandler callback). The Kafka client uses a similar strategy for >> producing messages by dynamically allocating a pool of ByteBuffers based on >> available memory. >> >> Also, I don't have much experience with this, but if we had a pool of >> reusable ByteBuffers, could we use direct allocation to get off-heap >> buffers? That seems like an interesting use case. >> >> -- >> Matt Sicker <boa...@gmail.com> >> >> >> -- > Matt Sicker <boa...@gmail.com> > -- Matt Sicker <boa...@gmail.com>