Releases: hhu-bsinfo/hadroNIO
Releases · hhu-bsinfo/hadroNIO
v0.3.3 - Minor release:
-Update dependencies (JUCX 1.13.0) -WrappingSocket: Log a warning instead of throwing an exception, when an unsupported method is called -example: Remove netty hello demo -example: Add gRPC demos (echo, benchmark and a simple key-value store with a YCSB-binding)
v0.3.2 - Minor release:
-Update dependencies (JUCX 1.12.1) -example: Add result to combiner before closing a channel to avoid lost results -example: Combine message count for multi-threaded benchmarks -example: Fix netty throughput benchmark hanging on last messages for odd message counts -example: Add option to pin threads in netty benchmarks -example: Adapt netty worker threads to connection count
v0.3.1 - Minor release:
example: Support multi-threading in netty latency benchmark example: Add option to write CSV formatted benchmark results to a file example: Add option to disable hadroNIO via an environment variable example: Add automated benchmark scripts
v0.3.0 - Third release:
-Refactor interfaces for UCX bindings -Improve error handling: Errors occuring inside a UCX binding are escalated to the Selector, which handles the errors or throws appropriate exceptions -SocketChannel: Fix flush not being handled properly by write (Fixes possible receive side overload, which can result in high memory usage by UCX) -SocketChannel: Reimplement gathering write (Buffers are now merged to drastically improve perfomance for small messages) -SocketChannel: Verify tags using checksums -SocketChannel: Fix possible hangup during tag exchange -example: Gather messages in throughput benchmarks (Maximum amount of gathered messages can be configured via a new parameter) -example: Support multi-threading in netty throughput benchmark -example: Add warmup to benchmarks -Add initial support for infinileap (not stable yet)
v0.2.0: Second release:
-SocketChannel: Use one worker per channel to avoid race conditions in UCX and events being detected by wrong selectors -Selector: Use busy polling to adapt to the new worker setup (temporary solution until epoll based polling is implemented) -Selector: Implement timeout -JucxWorker: Do not synchronize calls to progress() anymore, since it is no longer necessary -ServerSocketChannel: Fix possible NullPointerException, when calling accept() without any pending connection requests -Configuration: Update default configuration to use 8 MiB buffers with 64 KiB slices -Example: Implement test applications based on netty
v0.1.2 - Minor release:
-Selector: Fix select() possibly returning, although no keys have been selected -Selector: Fix wakeup() -Selector: Add debug log messages -JucxWorker: Synchronize calls to progress() as a workaround for crashes in progressWorkerNative(), when multiple threads operate on the same worker instance
v0.1.1 - Minor release:
-Implement Socket and ServerSocket as wrapper classes -> Calls to socket() are now working -SocketChannel: Do not report as connected until finishConnect() has been called (non-blocking mode only) -SocketChannel: Return -1 in read() if underlying channel implementation has lost connection -JucxSocketChannel: Close channel if remote endpoint becomes unreachable -Based on JUCX 1.11.2
v0.1.0 - First release:
-Acceleration for blocking and non-blocking socket channels is supported -Java 8+ is supported -Based on JUCX 1.11.0