This page describes our core software engineering values and explains why we are motivated to work on this project. Here is the TL;DR version:
We strive to
- Simulate practically useful and large engineering problems, with a
- Production-quality code that is documented, tested, extensible, and maintainable, with
- Economic use of hardware resources, even with a priori unknown and heterogeneous load distribution that dynamically changes during computation, and even on heterogeneous hardware whose components' performance dynamically changes during computation, featuring
- Asynchronous parallel communication by default, to free the computation from the undue influence of external factors, such as an artificially imposed clock or waiting for slower processing elements,
- Dynamic load balancing, to enable coping with a priori unknown load due to either hardware heterogeneities or adaptive algorithms, and
- Automatic object migration, to free the application of hopelessly intertwining physics, algorithms, and complex load balancing code, to ensure sustainability.
Our target machines are the largest distributed-memory computers in the world with potentially millions of compute cores. Due to unprecedented hardware complexity and features such as deep memory hierarchies and dynamic frequency scaling, we must assume a priori unknown and inhomogeneous computational load and performance among parts of the system that can also dynamically change in time. As large problems are partitioned into smaller chunks, information along partition boundaries, that exist on multiple processing elements, must be made consistent, which requires communication. To efficiently use resources our programming paradigm must allow asynchronous parallel execution and communication. Using asynchronous communication data is not transmitted at regular intervals and the transmitter and receiver do not have to be synchronized. Compared to synchronized (or blocking) communication, such asynchronous communication enables overlapping computation, communication, input, and output. However, asynchronous programming constitutes a major paradigm shift compared to the more traditional bulk-synchronous approach, widely applied for large-scale scientific computing, e.g., using message passing. We believe that the most economic utilization of future computer hardware can only be achieved with a paradigm shift from bulk-synchronous to fully asynchronous programming. For most computational codes this deeply affects the programming style as well as the optimal data layout and algorithm structure which therefore would require a complete rewrite.
Quinoa was started from scratch, instead of modifying an existing code, to allow full freedom in exercising the asynchronous paradigm. The code is built on the Charm++ runtime system and family of libraries. Instead of message passing, Charm++ is founded on the migratable-objects programming model and supported by an adaptive runtime system. In Charm++ data and work-units interact via asynchronous function calls enabling fully asynchronous programming. Asynchronous programming can be used to specify task-parallelism as well as data parallelism in a single application using a single abstraction. The interacting objects may reside on the same or on a networked compute node and may migrate from one to another during computation. Object migration is transparent to the application and carried out by the runtime system based on real-time load and hardware measurement, but if needed can be influenced by the application. Charm++ is mature, it has been actively developed since 1989, and is used by several large production codes. Read a one-page summary on the strengths of Charm++ at http:/
Nothing is more important than code that works as advertised with no surprises. We strive for writing testable code as well as writing tests that cover the code to the maximum degree possible. Code coverage, the percentage of lines of code tested compared to all the code, is quantified using two independent tools: codecov.io and gcov. Untested code is assumed to be incorrect until proven otherwise. Only with extensive positive and negative testing can developers and users be assured of correctness.
In Quinoa code correctness is verified and quantified using multiple levels of testing:
- Our random number generator test suite, RNGTest, subjects generators to stringent statistical tests which enables quantitative ranking of all generators available with respect to their quality and computational cost.
- Our unit test suite, UnitTest, capable of testing serial, synchronous (e.g., MPI) parallel, and asynchronous parallel (e.g., Charm++) functions with code coverage analysis is used to test the smallest units of code.
- Our regression test suite with code coverage analysis is used to test features larger than the smallest units, such as multiple algorithms coupled to solve a differential equation.
- A planned fuzz test suite will help prepare against invalid, unexpected, or random inputs.
Power consumption is a primary concern in the exascale era. Application performance must be measured within power constraints on increasingly complex and thus likely less reliable hardware. A computation must optimize and thus dynamically adapt to maximize performance constrained by power limits while being resilient against hardware failures. This requires revolutionary methods with a stronger-than-before integration among hardware features, system software, and the application.
Quinoa relies on the Charm++ runtime system. A central idea of Charm++ is to enable and facilitate overdecomposition: computation (data and work-units) is decomposed into a large number of logical units, usually more than the available processors. Overdecomposition enables the runtime system to dynamically adapt the computational load monitoring load-imbalance due to software (e.g., particle clustering, adaptive refinement) as well as hardware (e.g., dynamic processor frequency scaling). The runtime system also migrates data and work-units if it notices (via sensors, cache monitors, etc.) that a compute node is about to fail. If a node fails without a warning, the application can be restarted from a previously saved checkpoint. Since work decomposition and parallel programming are done without direct reference to physical processors, the application can be restarted using a number of processors different than that of the checkpoint was saved with. Resiliency is provided by the runtime system transparent to the application and can save millions of cycles since jobs have to be restarted less frequently due to hardware failure. Read more on power, reliability, and performance from the developers of Charm++.
Hardware complexity is increasing. Simultaneously satisfying all the different requirements enumerated here inevitably increases software complexity. Attempting to tackle every aspect by the application programmer, as is frequently done in a research context and sometimes also in production, does not scale as more features are added, i.e., not economical thus unsustainable. The combination of increasing hardware and software complexity leads to an unprecedented degree of specialization among the software components as well as their developers. Picking the right tool for the right job, components of complex software must be outsourced to subject-matter experts.
Quinoa's goal is to provide simulation software for scientific and engineering purposes. This involves numerically solving the differential and integral equations of mathematical physics. We cannot claim to be experts in all ingredients required, therefore highly specialized components, e.g., advanced computer science, such as load decomposition, dynamic load balancing, object migration, low level hardware-specific networking, parallel input and output, random number generation, hashing, etc., are outsourced to those who make a career out them. We subscribe to the proudly found elsewhere paradigm, instead of the not invented here stance. Accordingly, Quinoa uses a number of third-party libraries.
The ultimate measure of the value of a computer language is how it balances runtime performance and code complexity. A good language can do a lot for a designer and a programmer, as long as its strengths and limitations are clearly understood and respected.
Quinoa's main language of choice is C++ for the following reasons.
- It offers uncompromising performance.
- It is statically typed, resulting in earlier error detection, more reliable algorithms, and compiled code that executes quicker.
- It supports data abstraction and enables the representation of non-trivial concepts.
- It allows programming using more than one programming style, each to its best effect; as such it supports procedural, object oriented, generic, as well as many features of functional programming.
- It provides a portable standard library with guaranteed algorithmic complexity
Modern C++ provides great flexibility and enables the expert programmer to implement capability to simulate interesting (i.e., complex and practically useful) problems, using hardware resources efficiently, yielding production quality code that is extensible, maintainable, and thus sustainable.
Here are our coding guidelines, listed most important first:
- Correctness – Correct results, resilience against errors, fault tolerance.
- Performance – Maximize FLOPS, minimize communication.
- Power consumption – Minimize power required for given performance.
- Maintainability, easy to read and use – Maximize code-reuse.
- Easy to write – Optimize code for being read even at the expense of making code harder to write.
Some mechanism, preferably the runtime system with some help from the application, should monitor the first three aspects and dynamically influence the algorithm. This requires runtime introspection, which must be designed into the algorithms at the outset.
Quinoa uses the Charm++ runtime system capable of runtime introspection. Charm++ facilitates task-based parallelism, automatic load balancing through network-migration of objects, and enables coping with hardware heterogeneity and system failure.
Due to software complexity the most expensive resource in implementing, maintaining, and extending compute capability is the developer's time. In a successful project programmer productivity is highly regarded and actively maintained. Besides an inspiring and motivating culture, this involves the freedom to use the right software abstraction for the job at hand using the best and most versatile tools available. Only by using the latest and greatest tools are code reuse, extensibility, maintainability, and thus productivity maximized.
Quinoa is developed using the latest compiler technology and software engineering tools. We believe developers of a computational physics code must be skilled not only in physics and numerical methods but in the latest software development techniques as well.
User experience is the most important design goal of desirable software. However, this does not necessarily require a graphical user interface. Also, expert developers should be able to get started quickly, extend the code in a productive manner, and tailor functionality to their or their customer's needs.
User input in Quinoa is restricted to command-line arguments and simple-to-read text files. The input grammar is versatile and extensible. Parsing (but not the grammar) is outsourced to a library written by experts in that field. Error messages are friendly and often suggest a solution. Documentation of command-line and input file keywords is directly accessible at the user's fingertips, obtained via command-line arguments (
-H). New keywords, reflecting new features, are made impossible to add without proper in-code documentation. Simulation result output is highly customizable.
Undocumented code is considered legacy code and a major impediment to progress and productivity.
Quinoa is well-documented for developers, users, administrators as well as auditors. This includes documentation for
- software requirements, specification, design, implementation, and interfaces,
- verification and validation,
- user examples,
- source code control history,
- team collaboration documentation and archive,
- code correctness and quality,
- legal issues.
All documentation is accessible via a web browser featuring an expertly-designed search capability, with no large separate documents to open or print. Documentation is added via editing the source itself and looks great on any device, including figures and math.
We believe all the above are important in order to make this project fun to work on. Consequently, none of the above can be an afterthought: they must all be simultaneously considered at all times and at the outset.