March 2019 Monthly Update
Clojurists Together News
Hello CT Community -
The second month of updates from Neanderthal and Aleph is here.
Thanks for your continued support of Clojurists Together!
- Deep Learning from Scratch to GPU - 9 - The Activation and its Derivative
March 20, 2019 We implement the key part of the backward pass, the computation of the error of a layer. Along the way, we set up the infrastructure for the complete implementation of backpropagation.
- Deep Learning from Scratch to GPU - 10 - The Backward Pass (CUDA, OpenCL, Nvidia, AMD, Intel
March 25, 2019 We complete the basic implementation of the backward pass of backpropagation and gradient descent.
- Deep Learning from Scratch to GPU - 11 - A Simple Neural Network Inference API
March 28, 2019 The time is ripe for wrapping what we have built so far in a nice Neural Network API. After all, who would want to assemble networks by hand?
I've spent quite a lot of time working through comments for previous PRs and keeping those changes in a mergable state. At the end of the day a few PRs made their way to master and I hope to see more following.
- Issue with WebSocket per-message deflate extension handshaker turned out to be way more complex than I’ve expected. I had an idea to close the gap of inconsistent APIs on Netty's side: “Decouple WebSocket server extension handshaker from read I/O“. But as these changes didn't get through, I'm going to implement workaround in Aleph's codebase.
- Expose more information about handshake result both on the client and on the server, e.g. subprotocols and extensions #498.
- WebSocket client handshake processing flow changes, #498. This change effectively reverts one of my old PRs with regards to how aggregated HTTP request is process when doing WebSocket handshake. Previously I thought that this approach leads to memory leaks, which turned out not to be the case. New implementation should be more performance and obviously has less memory overhead.
- A few updates to WebSocket client force close after given timeout functionality submitted to Netty.
- I also spent tons of hours investigating performance of
netty/sinkin turms of latencies, throughput, backpressure. I was looking into Netty's
autoReadsemantic, how Aleph deals with it and if we can improve performane oding
readComplete. To make story short: we can. In cases when you send/recieve a lot of small chunks. I'm still not sure how to expose this in the framework tho'. Talking about small chunks of data... I did the same investigation for UDP.
- #497 Use
HttpUtilto manage "Content-Length" header, small internal improvement.
- Even more fixes for Keep-Alive connections management for Netty's examples, I think it makes sense to refactor part of the Aleph and move Keep-Alive logic into a separate channel handler so I would be easier to extend/fix (e.g. to close the connection after the error response).
- Nasus (static files server) got a few more features: CORS headers, symlics support, Basic Auth.
- HTTP client timeouts were reimplemented with
HashedWheelTimers that are designed specifically to handler tons of I/O events, #499. That's a huge performance improvement for projects with higher RPS rates. The reasons for that were described in #479, thanks to Alexander Yakushev for brining this up.
I've started this as an experiment.. but now part of client-side logic is implemented: connect/disconnect logic, publish messages to topics, acknowledgements flow for at least once QoS level. There're still a lot of things missing but end-to-end example already works. I hope I'll manage to pack those into a separate PR pretty soon.