Open Interface Enables Parallel Applications Debugging

By Michael Feldman

August 25, 2006

Like many companies who initially developed their offerings for the high performance computing market, Etnus LLC is finding that their TotalView debugger product has the potential for a much larger audience. Following the HPC trend in multi-core processors, multi-processor servers and cluster architectures into the commercial space, Etnus is demonstrating the competency of its debugger for a wide variety of multi-threaded software.

“HPC is our home base and core demographic,” says Chris Gottbrath, Product Manager for the TotalView debugger at Etnus LLC. But he notes that increasingly their customers come from the entertainment, financial, oil & gas and e-commerce industries. Companies in these industries are combining Linux and other open source software with the HPC competency of vendors like Etnus to build their solutions.

Gottbrath reports that Etnus has been helping to develop open interfaces for MPI (Message Passing Interface) to enable parallel application debugging. Today most supercomputers use MPI to support parallelization on distributed computing architectures, such as clusters. As clusters become widespread in enterprise deployments, more users will be relying on MPI to run their commercial applications. Gottbrath's highlights Etnus' collaboration with some of the MPICH and Open MPI developers to extend these interfaces for the next generation of the Message Passing Interface — MPI 2. HPCwire caught up with Gottbrath recently to discuss these latest developments.

HPCwire: Can you briefly describe how Etnus' TotalView debugger performs parallel debugging and talk about the importance of the MPI interface?

Gottbrath: TotalView performs parallel debugging by itself becoming a parallel application running alongside the user's application. TotalView begins by starting a lightweight (low resource usage) debugging agent on each node that is running part of the user's MPI program. These debugging agents — called tvdsvrs — can each use operating system level interfaces to interact with the users MPI tasks that are running alongside them on the same nodes. TotalView then controls the activity of all these lightweight debugging agents to provide the user with a comprehensive view into and fine level of control over the behavior of their MPI application that they are trying to debug.

In order to do this, one of the things that TotalView needs is a way to discover the identity and location of the potentially thousands of processes that make up an MPI job. The MPI process acquisition interface provides a way for the MPI itself to provide that information to the debugger.

Another interface that I will talk about is one that we use to obtain information about the state of the MPI message queues. Typically these queues are just used in a very transient way by the MPI implementations as they are processing messages. However when a message passing program fails to communicate correctly it often ends up leaving messages 'stuck' in the MPI message queues. This interface enables the debugger to give the user various high level and detailed views of these messages — and a quick way to diagnose an otherwise challenging set of problems.

Both of these interfaces were designed in collaboration with the MPI library development community and have been treated as open interfaces. Each interface has a reference implementation in the MPICH code base — so not only is the interface open but we can point MPI developers thinking about implementing the interfaces in their own code base to a full implementation that they can study. I think this has really fostered adoption of these interfaces both within the MPI library development community but also with a couple of other projects that have similar needs.

While I am talking about the existing interfaces I do want to place credit where credit is due. These interfaces are largely the work of Jim Cownie — who was at Etnus at the time, though is now at Intel. Jim worked closely with Rusty Lusk, one of the principle MPICH developers at Argonne on the original implementation. I've had the good fortune to be able to work with a wide range of other MPI developers as they work to incorporate these interfaces into their own implementations.

I am now leading a collaboration with the folks at MPICH and Open MPI to develop a new process acquisition interface that is capable of handling MPI 2 dynamic processes. We are quite consciously following the strategy of collaborative development, open interfaces, and open-source reference implementations that were so successful with these other two interfaces.

HPCwire: What specific kinds of information are available from the current MPICH and Open MPI interface that are being used by the TotalView debugger?

Gottbrath: The current MPI process acquisition interface tells TotalView the identity and location of the processes that make up an MPI job. One of the clever things about the interface is that it uses the fact that the debugger can read and write data from processes being debugged as the transport. What the MPI (or MPI runtime starter program) does is build up a table in memory that TotalView is able to look up like any other variable and read. Because MPI 1 has a very simple process lifetime — all MPI processes are created as a group and all of them are assumed to stay around till the whole program finishes — we only need a fairly light set of synchronization mechanisms. The debugger notifies the process that it is being debugged by setting a variable, the program starts the various MPI tasks in such a way that they won't run away, builds the table in memory and then notifies TotalView that the table is ready to be read, finally TotalView notifies the process that it is done reading the table and any remaining held tasks can be allowed to continue.

This is all done quite neatly with specially named functions and variables on the part of the MPI and the fundamental debugger operations on the part of TotalView.

The MPI message queue display interface provides the debugger with the envelope and contents of each message in the MPI send, receive or unexpected queues for each MPI process. Its implementation is a bit more sophisticated but operates on similar principles. Implementations take the form of helper libraries built by the MPI implementor to meet an API specified by Etnus. Each library encapsulates any implementation specific knowledge or techniques required to recover the message queue information out of a process. The helper library is loaded into the debugger which then provides low level memory read and write access to the MPI processes being debugged and makes API calls into the library as the user asks to look at the MPI message queues.

HPCwire: In your collaboration with the MPICH and Open MPI developers, what types of MPI interface enhancements are you most concerned about?

Gottbrath: My current work with the great folks in the Open MPI and MPICH teams focuses on capturing the good parts of the MPI process acquisition interface but extending the interface to allow the interface to be used in conjunction with MPI 2 dynamic processes. MPI 2 dynamic processes allow MPI programs to spawn new processes, join together previously separate MPI programs, and in certain circumstances split them back apart again. This makes for a much more complex world of MPI processes that the debugger needs to be made aware of. Synchronization is more complex and the performance-oriented nature of MPI means that we need to think very carefully about the costs of things like gathering together all the information about MPI processes in centralized locations.

My major concern as we design the interface is that like a good theory of the universe it be kept as simple as it can be, but no simpler. It needs to describe a complex situation, with MPI processes coming and going. It also needs to do this in a way that sacrifices little or no runtime performance. One of the major factors is that we want the MPI implementors to be comfortable shipping their libraries configured so that this information is gathered and made available to the user by default. Users will want to attach TotalView to their MPI jobs when they hang or crash and a nifty interface doesn't do anyone any good if it isn't turned on when the user sits down to troubleshoot their application.

HPCwire: What kinds of additional functionality would be available for TotalView users if you incorporated the MPI 2 interfaces that you are seeking?

Gottbrath: The principle advantage for users will be the ability to seamlessly debug MPI applications that take advantage of the new dynamic process features of MPI 2. MPI 2 was defined quite a while ago but only now are a number of different MPI libraries starting to support MPI 2 dynamic process operations. These are not something that every program is going to be immediately rewritten to start using, but they do open up significant new possibilities in the way that scientists and engineers can think about their MPI applications. Previously in order to use MPI the developer had to base their application on a fixed set of processes decided at launch time. With MPI 2 the set of processes that make up the MPI job can change in real time to respond to the demands of the task at hand.

—–

Chris Gottbrath

 

Chris Gottbrath is Product Manager for the TotalView Debugger at Etnus LLC. He is focused on making it easier for programmers, scientists and engineers to solve even the most complex bugs and get “back to work.” He has pursued this goal in a variety of customer-focused technical roles at Etnus over the last 6 years. Prior to that, as a graduate student of astrophysics at the University of Arizona in Tucson, he wrote cosmological simulations (with the occasional bug) using C and MPI on a small-scale Beowulf cluster.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion XL — were added to the benchmark suite as MLPerf continues Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing power it brings to artificial intelligence.  Nvidia's DGX Read more…

Call for Participation in Workshop on Potential NSF CISE Quantum Initiative

March 26, 2024

Editor’s Note: Next month there will be a workshop to discuss what a quantum initiative led by NSF’s Computer, Information Science and Engineering (CISE) directorate could entail. The details are posted below in a Ca Read more…

Waseda U. Researchers Reports New Quantum Algorithm for Speeding Optimization

March 25, 2024

Optimization problems cover a wide range of applications and are often cited as good candidates for quantum computing. However, the execution time for constrained combinatorial optimization applications on quantum device Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at the network layer threatens to make bigger and brawnier pro Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HBM3E memory as well as the the ability to train 1 trillion pa Read more…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HB Read more…

Nvidia Looks to Accelerate GenAI Adoption with NIM

March 19, 2024

Today at the GPU Technology Conference, Nvidia launched a new offering aimed at helping customers quickly deploy their generative AI applications in a secure, s Read more…

The Generative AI Future Is Now, Nvidia’s Huang Says

March 19, 2024

We are in the early days of a transformative shift in how business gets done thanks to the advent of generative AI, according to Nvidia CEO and cofounder Jensen Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Nvidia Showcases Quantum Cloud, Expanding Quantum Portfolio at GTC24

March 18, 2024

Nvidia’s barrage of quantum news at GTC24 this week includes new products, signature collaborations, and a new Nvidia Quantum Cloud for quantum developers. Wh Read more…

Alibaba Shuts Down its Quantum Computing Effort

November 30, 2023

In case you missed it, China’s e-commerce giant Alibaba has shut down its quantum computing research effort. It’s not entirely clear what drove the change. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

DoD Takes a Long View of Quantum Computing

December 19, 2023

Given the large sums tied to expensive weapon systems – think $100-million-plus per F-35 fighter – it’s easy to forget the U.S. Department of Defense is a Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Leading Solution Providers

Contributors

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Google Introduces ‘Hypercomputer’ to Its AI Infrastructure

December 11, 2023

Google ran out of monikers to describe its new AI system released on December 7. Supercomputer perhaps wasn't an apt description, so it settled on Hypercomputer Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Intel Won’t Have a Xeon Max Chip with New Emerald Rapids CPU

December 14, 2023

As expected, Intel officially announced its 5th generation Xeon server chips codenamed Emerald Rapids at an event in New York City, where the focus was really o Read more…

IBM Quantum Summit: Two New QPUs, Upgraded Qiskit, 10-year Roadmap and More

December 4, 2023

IBM kicks off its annual Quantum Summit today and will announce a broad range of advances including its much-anticipated 1121-qubit Condor QPU, a smaller 133-qu Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire