Blogs

Paris Carbone, Asterios Katsifodimos, Stephan Ewen, Volker Markl, Seif Haridi, Kostas Tzoumas, presented in the world’s most confusing game of asterisks I’ve seen so far. Bottom Line Up Front I rate this 3.9. Flink is an academic attempt at replacing Spark. I haven’t figured out why. I guess just even higher level/more optimizations? Or maybe I’m late to the party and most of these have spilled into Spark? Flink programs compute both early and approximate and delayed and accurate results in the same operation....

Jesse Dodge, Taylor Prewitt, Remi Tachet Des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, NOah A. Smith, Nicole DeCario, Will Buchanan Bottom Line Up Front I rate this 6.6. Two papers wearing a trenchcoat. The first half is about how much electricity machine learning models use, and lots of experiments and high quality data are used to work this out. The second is two half-baked ideas about what to do with this information, with a few strange figures spit out hinging on assumptions that have nothing to do with either machine learning algorithms or reality....

Robert Metzger - GOTO 2019 What is Flink? Low latency, high throughput, stateful, distributed stream processing framework. Stateful Computations over Data Streams You can use this for batch processing, static or historic data in a fast way. Or, you are processing realtime data, processing a stream of data and updating your model of the world. Or, event-driven applications. 3 Use Cases Streaming ETL Traditionally, ETL is a periodic job fired off by cron....

Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach Mike Burrows, Tushar Chandra, Andrew Fikes, Robert E. Gruber Bottom Line Up Front I rate this Summary Introduction Distributed storage system. PB scale. Very applicable, scalable, performant, and available. 60 projects using it. BigTable is like a database but not relational. Everything is a string, clients need to cope with it. Data Model Sparse, distributed, persistent multi-dimensional sorted map....

Udit Gupta et al. in a wonky curvy font. Bottom Line Up Front I rate this 5.1 New model is proposed. Multiple different ways to use it depending on what the type of hardware is. Goal is for designers themselves to build hardware that minimizes life-cycle impact, not just operational. Summary Intro We want to reduce consumption for the environment, we want to increase consumption for all the new applications and cloud-scale fun....

Hashmi, Akbar, Adve, Moo, Ding Bottom Line Up Front Very dense, even for a lit review. Well structured, and will be useful as a reference once I start digging into a problem. Only obvious extension I can think of is looking for adjacent problems in other fields. I know bin packing and estimating state from measurements isn’t unique to radars, would be nifty to see what is going on outside of the ECE department....

Kosuru, Qu, Ding, Moo. Bottom Line Up Front RL agent is forced to pick one of 4 schedulers. Usually hones in on the best one. Summary Introduction Radars can have many tasks. Tasks have priority $p$, times $t_{start}$,$t_{dwell}$. Windows have a length $L$. Time frame lasts $t_{total}$. Radar resource management “RRM” tries to maximize utilization of $L$ by dropping some tasks during overloading situations. Branch and Bound method is great but computationally expensive....

Bottom Line Up Front Acronyms MFR: multi-function radar RRM: radar resource management EST: earliest-start-time (scheduling method) ED: earliest-deadline (secheduling method) DSS: dual side scheduling NCT: nearest closer time RSST: Random shifted start time Summary Instead of jamming tasks left, toss out some random points and scrunch them together about them. Fast enough to calculate, drop fewer and cost’s less after checking a bunch for one that costs less. Intro Phased array radars have multiple functions, which need different tasks completed....

A paper by Udit Gupta about the environmental impact of computing systems. Bottom Line Up Front I rate this 4.1. Making computers is worse than using computers. Comprehensive overview of all of the different sources of carbon in computing, very citation dense, and probably best used as a jumping-off point to go in-depth into any one of the paragraphs on this paper. Summary Information and communication technology includes consumer devices, networking tech, and data centers....

Diego Ongaro Bottom Line Up Front I rate this 7.7. Raft is easier than Paxos to understand. Here’s how Raft works. Logs, pick a leader, commit once majority has written to disk. Elections rely on majorities. Write the correct things to disk, use random delays to break ties. Makes me want to read about Chubby. Summary Consensus algorithms allow a collection of machines to work as a coherent group that can survive some failures....