Section: Research Program

Structuring Applications

Computing on different scales is a challenge under constant development that, almost by definition, will always try to reach the edge of what is possible at any given moment in time: in terms of the scale of the applications under consideration, in terms of the efficiency of implementations and in what concerns the optimized utilization of the resources that modern platforms provide or require. The complexity of all these aspects is currently increasing rapidly:

Diversity of platforms.

Design of processing hardware is diverging in many different directions. Nowadays we have SIMD registers inside processors, on-chip or off-chip accelerators (GPU, FPGA, vector-units), multi-cores and hyperthreading, multi-socket architectures, clusters, grids, clouds... The classical monolithic architecture of one-algorithm/one-implementation that solves a problem is obsolete in many cases. Algorithms (and the software that implements them) must deal with this variety of execution platforms robustly.

As we know, the “free lunch” for sequential algorithms provided by the increase of processor frequencies is over, we have to go parallel. But the “free lunch” is also over for many automatic or implicit adaption strategies between codes and platforms: e.g the best cache strategies can't help applications that accesses memory randomly, or algorithms written for “simple” CPU (von Neumann model) have to be adapted substantially to run efficiently on vector units.

The communication bottleneck.

Communication and processing capacities evolve at a different pace, thus the communication bottleneck is always narrowing. An efficient data management is becoming more and more crucial.

Not many implicit data models have yet found their place in the HPC domain, because of a simple observation: latency issues easily kill the performance of such tools. In the best case, they will be able to hide latency by doing some intelligent caching and delayed updating. But they can never hide the bottleneck for bandwidth.

HPC was previously able to cope with the communication bottleneck by using an explicit model of communication, namely MPI. It has the advantage of imposing explicit points in code where some guarantees about the state of data can be given. It has the clear disadvantage that coherence of data between different participants is difficult to manage and is completely left to the programmer.

Here, our approach is and will be to timely request explicit actions (like MPI) that mark the availability of (or need for) data. Such explicit actions ease the coordination between tasks (coherence management) and allow the platform underneath the program to perform a pro-active resource management.

Models of interdependence and consistency.

Interdependence of data between different tasks of an application and components of hardware will be crucial to ensure that developments will possibly scale on the ever diverging architectures. We have up to now presented such models (PRO, DHO, ORWL) and their implementations, and proved their validity for the context of SPMD-type algorithms.

Over the next years we will have to enlarge the spectrum of their application. On the algorithm side we will have to move to heterogeneous computations combining different types of tasks in one application. For the architectures we will have to take into account the fact of increased heterogeneity, processors of different speed, multi-cores, accelerators (FPU, GPU, vector units), communication links of different bandwidth and latency, memory and generally storage capacity of different size, speed and access characteristics. First implementations using ORWL in that context look particularly promising.

The models themselves will have to evolve to be better suited for more types of applications, such that they allow for a more fine-grained partial locking and access of objects. They should handle e.g collaborative editing or the modification of just some fields in a data structure. This work has already started with DHO which allows the locking of data ranges inside an object. But a more structured approach would certainly be necessary here to be usable more comfortably in applications.

Frequent IO.

A complete parallel application includes I/O of massive data, at an increasing frequency. In addition to applicative input and output data flow, I/O is used for checkpointing or to store traces of execution. These then can be used to restart in case of failure (hardware or software) or for a post-mortem analysis of a chain of computations that led to catastrophic actions (for example in finance or in industrial system control). The difficulty of frequent I/O is more pronounced on hierarchical parallel architectures that include accelerators with local memory.

I/O has to be included in the design of parallel programming models and tools. ORWL will be enriched with such tools and functionalities, in order to ease the modeling and development of parallel applications that include data IO, and to exploit most of the performance potential of parallel and distributed architectures.

Algorithmic paradigms.

Concerning asynchronous algorithms, we have developed several versions of implementations, allowing us to precisely study the impact of our design choices. However, we are still convinced that improvements are possible in order to extend its application domain, especially concerning the detection of global convergence and the control of asynchronism. We are currently working on the design of a generic and non-intrusive way of implementing such a procedure in any parallel iterative algorithm.

Also, we would like to compare other variants of asynchronous algorithms, such as waveform relaxations. Here, computations are not performed for each time step of the simulation but for an entire time interval. Then, the evolution of the elements at the frontiers between the domain that are associated to the processors are exchanged asynchronously. Although we have already studied such schemes in the past, we would like to see how they will behave on recent architectures, and how the models and software for data consistency mentioned above can be helpful in that context.

Cost models and accelerators.

We have already designed some models that relate computation power and energy consumption. Our next goal is to design and implement an auto-tuning system that controls the application according to user defined optimization criteria (computation and/or energy performance). This implies the insertion of multi-schemes and/or multi-kernels into the application such that it will able to adapt its behavior to the requirements.