28 March 2023 in IEEE Design & Test
TestRIG (Testing with Random Instruction Generation) is a testing framework for RISC-V implementations. The RISC-V community has standardized a formal model of the architecture in the Sail language 1 , giving a human-readable specification that can also be used for simulation and verification. The Sail language is designed for this purpose by allowing instruction semantics to be described conveniently (for example, by supporting variable bit-widths). Ideally, a RISC-V implementor could formally prove equivalence between their implementation and the Sail model, but proof tools are not yet sufficiently automated to be routinely used on the whole-processor level. They instead focus on equivalence between combinational logic functions, with verification of a full out-of-order microarchitecture remaining an open problem. As a pragmatic compromise, we use TestRIG to check equivalence between the model and an implementation by generating random instruction sequences, executing the same sequences on the model and the implementation under test, and comparing execution traces (tandem execution). This approach does not prove equivalence but can demonstrate divergence, and is usable in all stages of development.
20 June 2023 in ACM Transactions on Parallel Computing
HPC clusters have become ever more expensive, both in terms of capital cost and energy consumption; some estimates suggest that competitive installations at the end of the next decade will require their own power station. One way around this looming problem is to design bespoke computing engines, but while the performance benefits are good, the design costs are huge and cannot easily be amortized. Partially Ordered Event Triggered System (POETS)—the focus of this article—seeks to exploit a middle way: The architecture is tuned to a specific algorithmic pattern but, within that constraint, is fully programmable. POETS software is quasi-imperative: The user defines a set of sequential event handlers, defines the topology of a (typically large) concurrent ensemble of these, and lets them interact. The “solution” may be exfiltrated from the emergent behaviour of the ensemble. In this article, we describe (briefly) the architecture, and an example computational chemistry application, dissipative particle dynamics (DPD). The DPD algorithm is traditionally implemented using parallel computational techniques, but we re-cast it as a concurrent compute problem that is then ideally suited to POETS. Our prototype system is realised on a cluster of 48 FPGAs providing 50K concurrent hardware threads, and we report performance speedups of over two orders of magnitude better than a single thread baseline comparator and scaling behaviour that is almost constant. The results are validated against a “conventional” implementation.
1 March 2022 in IET Computers & Digital Techniques
One of the key problems in designing and implementing graph analysis algorithms for distributed platforms is to find an optimal way of managing communication flows in the massively parallel processing network. Message‐passing and global synchronization are powerful abstractions in this regard, especially when used in combination. This paper studies the use of a hardware‐implemented refutable global barrier as a design optimization technique aimed at unifying these abstractions at the API level. The paper explores the trade‐offs between the related overheads and performance factors on a message‐passing prototype machine with 49,152 RISC‐V threads distributed over 48 FPGAs (called the Partially Ordered Event‐Triggered Systems platform). Our experiments show that some graph applications favour synchronized communication, but the effect is hard to predict in general because of the interplay between multiple hardware and software factors. A classifier model is therefore proposed and implemented to perform such a prediction based on the application graph topology parameters: graph diameter, degree of connectivity, and reconvergence metric. The presented experimental results demonstrate that the correct choice of communication mode, granted by the new model‐driven approach, helps to achieve 3.22 times faster computation time on average compared to the baseline platform operation.