January 11, 2020

237 words 2 mins read

Taking Full Advantage of NVME SSDs by Optimizing Page Cache Performance

Taking Full Advantage of NVME SSDs by Optimizing Page Cache Performance

Modern SSDs allow transfer rates via NVME of 2-3 Gbytes per second. One can hook up multiple of these amazing drives to a system but sadly most enterprise servers in their default configuration will t …
Talk Title Taking Full Advantage of NVME SSDs by Optimizing Page Cache Performance
Speakers Christoph H Lameter (Universalist specializing in Computer Science, None)
Conference Open Source Summit + ELC Europe
Conf Tag
Location Lyon, France
Date Oct 27-Nov 1, 2019
URL Talk Page
Slides Talk Slides
Video

Modern SSDs allow transfer rates via NVME of 2-3 Gbytes per second. One can hook up multiple of these amazing drives to a system but sadly most enterprise servers in their default configuration will top out at 3GB/sec and the aggregate performance of multiple of these drives (f.e. via a RAID0 configuration) cannot be effectively utilized. This is a particular problem with Intel processors due to their small 4K page size.The talk will investigate various ways to deal with this problem and show the advantages and limitations of each approach.A. Exploiting NUMA affinities to significantly increase performanceB. Direct I/OC. Huge pagesD. Various knobs in the Kernel configuration that affect performanceThere are ongoing efforts to make it easier to get full performance without using any of these techniques by allowing huge page support in the Linux page cache.In conclusion we will mention current issues to be aware of in current Enterprise distributions that affect performance.

performance exploit intel
comments powered by Disqus
Lightning Talk: Telemetry Aware Scheduling: Let the (Scheduling) Magic Unfold!!

Lightning Talk: Telemetry Aware Scheduling: Let the (Scheduling) Magic Unfold!!

January 7, 2020

With containers becoming the desired deployment model for a wide range of workloads such as Big Data, IoT, storage, AI, ML, SDN and NFV solutions, Container orchestration tools like Kubernetes, need t …
Practicing data science: A collection of case studies

Practicing data science: A collection of case studies

January 7, 2020

Rosaria Silipo shares a collection of past data science projects. While the structure is often similardata collection, data transformation, model training, deploymenteach required its own special trick, whether a change in perspective or a particular technique to deal with special case and special business questions.
nGraph: Unlocking next-generation performance with deep learning compilers

nGraph: Unlocking next-generation performance with deep learning compilers

January 4, 2020

The rapid growth of deep learning in demanding large-scale real-world applications has led to a rapid increase in demand for high-performance training and inference solutions. Adam Straw, Adam Procter, and Robert Earhart offer a comprehensive overview of Intel's nGraph deep learning compiler.
An active learning framework to optimize training of deep models with human in the loop

An active learning framework to optimize training of deep models with human in the loop

January 3, 2020

Humayun Irshad offers an overview of an active learning framework that uses a crowdsourcing approach to solve parking sign recognitiona real-world problem in transportation and autonomous driving for which a large amount of unlabeled data is available. The solution generates an accurate model, quickly and cost-effectively, despite the unevenness of the data.
Leveraging AI for social good

Leveraging AI for social good

December 30, 2019

The hardware, software, and algorithms that automatically tag our images or recommend the next book to read can also improve medical diagnosis and protect our natural resources. Jack Dashwood and Anna Bethke discuss a variety of technical projects at Intel that have enabled social good organizations and provide guidance on creating or engaging in these types of projects.
Turn devices into data scientistsat the edge

Turn devices into data scientistsat the edge

December 28, 2019

Todays approach to processing streaming data is based on legacy big-data centric architectures, the cloud, and the assumption that organizations have access to data scientists to make sense of it allleaving organizations increasingly overwhelmed. Simon Crosby shares a new architecture for edge intelligence that turns this thinking on its head.