> For the complete documentation index, see [llms.txt](https://xzhu0027.gitbook.io/blog/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://xzhu0027.gitbook.io/blog/ml-for-systems/ml-sys-index/short-summaries.md).

# Short Summaries

### [Neural Adaptive Content-aware Internet Video Delivery](https://www.usenix.org/system/files/osdi18-yeo.pdf) - Yeo et al., OSDI' 18

The key idea of this paper is to use **super-resolution**, a technique that uses DNN to recover a high-resolution image from lower resolution images, on top of ABR to enhance client-side video quality. It will train a DNN for each video offline, exploiting DNN's inherent overfitting property to guarantee reliable and superior performance. 

When a client requests a video from CDN server, the server provides the DNN corresponding to the video. The client then applies the DNN to the received low-quality video chunks by utilizing its own computing power. 

### [**Neural-Enhanced Live Streaming: Improving Live Video Ingest via Online Learning**](http://ina.kaist.ac.kr/~livenas/livenas_sigcomm2020.pdf) - Kim et al., SIGCOMM' 20

![](https://1271107977-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-LkwjM9iFxyVX0PINZ87%2F-MUG3CBS6-9seoK-kVwK%2F-MUG4zn5MqM26V8Y4Zbd%2FScreen%20Shot%202021-02-23%20at%202.52.07%20PM.png?alt=media\&token=87a7a91b-0f2e-4c5e-8bb5-354518662ec5)

The key observation is that the stream quality is fundamentally constrained by the streamer's uplink bandwidth and its computational capacity. Extending the idea of the above paper, this paper presents **LiveNAS**, a neural-enhanced live video streaming system, which breaks the strong dependency between the quality of live video and the ingest client's bandwidth.

Since it focuses on live streaming applications, pre-trained networks are not possible. Instead, LiveNAS trains super-resolution DNNs via online learning. Besides the encoded video, the client transmits small patches of high-quality raw frames which will serve as ground truth for training. Note that even a fraction of ground truth labels can provide substantial training gains because of the video's temporal redundancy.

### [**LinnOS: Predictability on Unpredictable Flash Storage with a Light Neural Network**](https://www.usenix.org/system/files/osdi20-hao.pdf) - Hao et al, OSDI' 20

![The request will be sent to a replica if the model predicts the response will be "slow"](https://1271107977-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-LkwjM9iFxyVX0PINZ87%2F-MUG3CBS6-9seoK-kVwK%2F-MUG9Ys0E374l9Fme3N5%2FScreen%20Shot%202021-02-23%20at%203.12.00%20PM.png?alt=media\&token=ac8f642c-3d8f-4e04-84d2-e52896daa587)

As SSD's internal complexity continues to grow, achieving highly predictable latency on modern flash devices is very challenging. This paper tries to use a simple neural network to learn the device behavior at per-I/O scale. 

The key characteristic/design decisions of LinnOS are 1) it converts the hard latency inference problem into a simple binary inference("fast" or "slow" speed). and 2) it makes binary inference on every incoming I/O with a light neural network in a black-box manner. 

Some details about the model:

* The model is a fully-connected neural network with only three layers. With several optimizations, it can achieve 4-6us inference speed
* The input to the model is the number of pending I/Os of that device +  the latency of N(e.g., 4) most-recently completed I/Os + the number of pending I/Os at the time when each of the R completed I/Os arrived
* To train the model, LinnOS uses the current live workload that the SSD is serving. Each traced I/Os is labeled as "fast" or "slow". 

### [**Interpreting Deep Learning-Based Networking Systems**](https://dl.acm.org/doi/10.1145/3387514.3405859) - Meng et al., SIGCOMM' 20

![](https://1271107977-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-LkwjM9iFxyVX0PINZ87%2F-MUGAJ0xije4p4paj8CK%2F-MUGCLbx4Ymw77mWtUKv%2FScreen%20Shot%202021-02-23%20at%203.24.21%20PM.png?alt=media\&token=bc926a98-c0f0-42fd-b90d-b04183795137)

Deep learning-based networked systems (e.g., [Pensieve](http://web.mit.edu/pensieve/) and [Decima](https://web.mit.edu/decima/content/sigcomm-2019.pdf)). treat DNNs as black-boxes, which makes them hard to debug, deploy, and adjust. Metis is a framework that provides interpretability of these systems and the goal is to interpret DL-based networked systems with human-readable control policies. It adopts a decision tree conversion method for local systems(i.e., systems that collect local information and make decisions for one instance only.) and a hypergraph conversion method for global systems(i.e., systems that aggregate information across the network and make global planning for multiple instances.)