PADDLE: Performance Analysis using a Data-Driven Learning Environment


Authors: Jayaraman Thiagarajan, Rushil Anirdh, Bhavya Kaikhura, Nikhil Jain, Tanzima Islam, Abhinav Bhatele, Jae-Seung Yeom, Todd Gamblin
Venue: IEEE International Parallel and Distributed Processing Symposium (IPDPS)

 In the scope of HPC, machine learning is gaining increased traction to add in performance analysis and tuning. However, this approach includes a pipeline of data collection, data pre-processing, various machine learning algorithm testing, tuning, and then finally trying to understand the model. The paper states that while this process is repetitive, rarely can insights be reused from one domain to another. To address this void, the propose PADDLE.

PADDLE has three key steps: deep feature extraction, model design, and visualization. The first step allows users to throw extensive amounts of data at the problem, and an automated solutions determines the key inputs, mapping them to a new feature space. The next step in paddle automatically tests a number of machine learning algorithms, selecting one which minimizes complexity and maximizes accuracy. Finally, the third step provides a visualization which provides insights as to the key inputs into the model, as well as showing relations between various sources.

PADDLE's generalized approach is tested on several domain specific areas, finding PADDLE to out-perform tailored approaches in several cases. My personal take away from this paper is that deep feature extraction seems to be a key component of what makes the pipeline successful, and I look forward to doing more reading in this specific domain. The visualization is a nice addition, but again, seems to be rooted in the feature extraction. The model design, or central component, is rather light in the text in terms of implementation. They use a notion of bias and variance in analyzing various  models, but again, note that much of the model success is tied to feature extraction. Overall, good insights into how domain specific problems can be generalized. Nevertheless, expertise will always be required in extracting features and selecting the right training data.

Full Paper

Comments

Post a Comment

Popular posts from this blog

Fundamental Latency Trade-offs in Architecting DRAM Caches (Alloy Cache)

Authors: Moinuddin K. Qureshi and Gabriel H. Loh Venue:    MICRO 2012 The authors present a new architecture for DRAM cache which targets optimizing latency rather than hit rate. The authors show that for every optimization made to improve hit rate, one must also consider the added cache latency, thus resulting in a new "break even hit rate". The authors compare to three key baselines: an infeasibly large SRAM-tag array, the prior "LH-Cache", and a ideal latency-optimized cache. They find that "unoptimizations" improve latency, while minimally impacting hit-rate. Their final result is a cache which boosts performance 35% improvement over the baseline, and nearly 20% better than the previous LH-cache. The specific contributions are: Moving to a direct mapped cache provides significant improvement (by reducing latency) Combining tag-store and data-store into a single entity improves performance 21% Implementing a small per-core predictor enhances per...
Read more

ZCOMP: Reducing DNN Cross-Layer Memory Footprint Using Vector Extensions

Authors: Berkin Akin, Zeshan A. Christi, and Alaa R. Alameldeen Venue: MICRO, 2019 With accelerators dominating the deep learning space in architecture conferences, this paper stands out as it focuses on reducing DNN inference/training overhead while using a CPU. An obvious question is, with the wide-spread use of GPUs/TPUs for deep learning, why should we focus on optimizing CPUs for deep learning. In a recent paper published at HPCA 2019, Facebook claims that CPUs are preferred for applications where a tight integration is required between DNN and non-DNN tasks. Also, Intel's recent AVX512 has specialized support for DNNs in the form of new instructions called Vector Neural Network Instructions (VNNI). Broadly, optimizing DNNs can be viewed from two different perspectives; computation, and communication. This paper targets reducing communication overhead, more specifically, activation or feature-map communication overhead by compressing them. Compressing activations/weights...
Read more

AutoFDO: Automatic Feedback-Directed Optimization for Warehouse-Scale Applications

Authors: Dehao Chen, David Xinliang Li, Tipp Moseley Venue:    CGO 2016 This paper presents AutoFDO, a system used which profiles warehouse-scale applications, and applies feedback to the compilation for the next release. AutoFDO works by profiling and storing profiles in an aggregate database, annotating the profiles via an intermediate representation, and finally providing feedback. On average, this technique boosts performance by around 10% and works well even with stale releases. For profiles to be useful, they must be in an intermediate representation. This is built from binary-level profile, which uses LBR to map instruction frequencies. By using these frequencies, program counters, and the source, a source profile can be generated. The source profile can be used to build an annotated call-graph with edge frequencies, with some inaccuracy. These annotated call-graphs can then be used as feedback to the compiler. Because of the size of the applications and scale of d...
Read more