UCX Hackton 2016

Meeting notes

Slides

Registration

Registration for the event

Dates

August 9-12

Location

• ARM, Austin
• Address: 1, 5707 Southwest Pkwy #100, Austin, TX 78735
• Google maps: click here
• Main Entrance:
• Parking Arrangements: a) You can park on the street b) Covered parking on the left (please make sure that the spot is not registered):

Agenda

We are planning to have the annual UCX meeting in spring of 2016. This page will track topics on the agenda and details of the meeting (TBD).

List of proposed topics:

• Invited talks

  • 2 confirmed talks ~45 each one. Should go as the first talk of the day.

• Intro to UCX

  • Project overview (4h, must of on the first day)

• UCP API aspects:

  • Non-contiguous data types - 2h (moderate priority)
  • UCP active messages - 2h (moderate priority)
  • Client/server connection establishment - 1h (moderate priority)
  • Error handling and fault tolerance - 1h (moderate priority)
  • Finalize UCP API - 2h (high priority, one of the last topic to finalize the meeting)

• UCP internals:

  • Ordering with tag matching over loose ordered transport(s) - 1h (high priorioty)
  • Multi-transport, multi-rail - 1h (high priority)
  • Interrupt driven progress - 1h (moderate priority)
  • ucp_progress (calling in other functions, multiple entrance, etc.) - 1h (high priority)

• GPU integration - 3h

• Results and conclusions

  • A comparison of UCX and libfabrics in terms of functionality and abstraction level - 1h

Schedule

Additional topics:

• Thread safety - Xin Zhao
• Case study: OpenSHMEM over UCX with uGNI - Matthew Baker
• OpenMPI integration with UCX
• Groups

Date	Time	Topic	Speaker
Aug-9	9:00	Registration	ARM Visitor Office
	9:30	Opening Talk	Pavel Shamis
	9:45	"Future Technologies"	Steve Poole
	10:30	UCX Introduction - Architecture Overview	Pavel Shamis
	11:15	Break
	11:25	UCX Introduction - UCP	Pavel Shamis
	12:15	Lunch (provided by ARM)
	1:00	UCX Introduction - UCT	Yossi Itigin
	1:45	UCX Introduction - UCS	Yossi Itigin
	2:30	Break
	3:00	"UCX support in MPICH" This talk will give a short introduction into the new CH4 layer in MPICH and the UCX implementation.	Kenneth Raffenetti and Lena Oden
	3:30	Case study: OpenSHMEM over UCX with uGNI	Matthew Baker
	4:00	OpenMPI integration with UCX	TBD
	4:30	UCP Internals - ucp_progress (calling in other functions, multiple entrance, etc...)	Lena Oden
	5:30	Break for the day
Aug-10	9:00	Coffee
	9:20	Highlights from prior day- intro theme of the day	Pavel Shamis
	9:30	"Supporting GPU Acceleration in Network" NVIDIA GPUs have been evolving and improving as accelerated computing units over the last 10 years, leading to their adoption in areas as diverse as robotics, autonomous driving, medical imaging, seismic analysis, machine learning and supercomputers. The explosion of innovation in deep learning on these GPUs in the last 3 years is now driving their adoption even faster. It is clear now that Exascale will happen through accelerated computing, and is much less than another 10 years away. It is essential and urgent that together we deliver a well-designed distributed system that enables well-orchestrated interaction between asynchronous compute tasks on the GPU, and data movement operations across the memories of multi-node systems. In this way we permit the strong scaling across these machines necessary to reach the goal of Exascale	Sylvain Jeaugey
	10:15	Break
	10:30	UCP Internals - Interrupt driven progress	Alex Margolin
	11:30	UCP Internals - Multi transport, multi rail	Yossi Itigin
	12:30	Lunch (provided by ARM)
	1:30	"UCX over Infiniband: Performance analysis and sources of overhead " This talk gives a deep analysis of UCP and UCT performance differences and analysis their sources. We identify some overheads that can be avoided to reach a better performance	Lena Oden, Nikela Papodopulus
	2:00	Thread Safety	Xin Zhao
	3:00	GPU Integration	Serguei Sagalovitch, Sylvain Jeaugey
	4:00	Break
	4:30	GPU Integration - continued	Serguei Sagalovitch, Sylvain Jeaugey
	5:30	Break
	6:30	Social gather event at Doc's backyard bar and Grill - 3 miles from ARM's office. Google map directions from ARM's office link here	Doc's Backyard Bar & Grill, 5207 Brodie Lane, Sunset Valley
Aug-11	9:00	Coffee
	9:20	Highlights from prior day- intro theme of day	Pavel Shamis
	9:30	"A System Software Approach for Enabling Integrated HPC and Big Data Applications" The notion that one operating system or a single unified software stack will support the emerging and future needs of the HPC and Big Data application communities is unrealistic. There are many technical and non-technical reasons why functional partitioning through specialized software stacks will continue to persist. Rather than pursuing a single software stack that satisfies a diverse and competing set of requirements, approaches that enable the use and integration of multiple software stacks should be pursued. This talk describes the challenges that motivate the need to support multiple concurrent software stacks for enabling application composition, more complex application workflows, and a potentially richer set of usage models for extreme-scale HPC systems. We describe the operating system infrastructure for supporting multiple concurrent software stacks that is being developed in the Hobbes OS project and discuss issues, challenges, and potential approaches for enabling integrated HPC and Big Data applications on extreme-scale computing systems.	Barney Maccabe, Oak Ridge National Laboratory (in collaboration with Ron Brightwell & Kevin Pedretti, Sandia National Laboratories and David Bernholdt, Oak Ridge National Laboratory)
	10:15	Break
	10:30	UCP API - Non contiguous data types	Richard Graham
	12:30	Lunch (provided by ARM)
	1:30	UCP API - UCP Active messages	Aurelien Bouteller
	2:30	Tag matching in UCT	Yossi Itigin
	3:30	Break
	4:00	UCP API - Client / Server connection establishment	Alex Margolin
	5:00	UCP API - Error handling and fault tolerance	Aurelien Bouteiller
	6:00	Break for the day
Aug-12	9:00	Coffee
	9:20	Highlights from prior day - intro theme of day	Pavel Shamis
	9:30	"Next Generation of Co-Processors Emerges – In-Network Computing". The latest revolution in HPC is the move to a co-design architecture, a collaborative effort among industry, academia, and manufacturers to reach Exascale performance by taking a holistic system-level approach to fundamental performance improvements. Co-design architecture exploits system efficiency and optimizes performance by creating synergies between the hardware and the software. Co-design recognizes that the CPU has reached the limits of its scalability, and offers an intelligent network as the new “co-processor” to share the responsibility for handling and accelerating application workloads. By placing data-related algorithms on an intelligent network, we can dramatically improve the data center and applications performance.	Gilad Shainer (Mellanox)
	10:15	Break
	10:30	Finalize UCP API	Pavel Shamis
	11:30	A comparison of UCX and libfabrics in terms of functionality and abstraction level	Howard Pritchard
	10:30	Parallel session - UCX board meeting
	1:30	Results and conclusions	Yossi Itigin