(D) Die nachfolgende Liste ist nicht vollständig. Es gibt immer studentische Arbeiten in verschiedene Forschungsprojekte bei uns am Lehrstuhl. Bitte bei Interesse uns eine Email schreiben (z.B. an Dai Yang oder PD Weidendorfer oder Vladimir Podolskiy).
(GB) The following list is not exhaustive. There is always some student work to be done in various research projects. You can send an email (e.g. to Dai Yang or PD Weidendorfer or Vladimir Podolskiy), asking for currently available topic.
Implementation and Evaluation of MLEM algorithm on Intel Xeon Phi Knights Landing (KNL) Processor
In a current project the Chair for Computer Architecture analyzes modern HPC system with heterogeneous architectures towards exascale computing. Real-world applications which represent a class of typical HPC problems are an important element. One example is the maximum likelihood expectation maximization (MLEM) algorithm [KWS+09], which is used for image reconstruction in positron emission tomography (PET). PET visualizes functional processes by measuring the distribution of a tracer of radioisotopes injected into a subjects’s body. Clinical PET scanners for example assist in tumor diagnosis. PET research currently focuses on improving spatial resolution and sensitivity of the technique. Our research is done on small animal PET scanners for preclinical stuies in cooperation with the Medical Institute Rechts der Isar (MRI). The MLEM algorithm is based on sparse matrix vector multiplication (SpMV). The efficient usage of heterogeneous systems with accelerator cards such as Intel Xeon Phi is still an open challenge. We have already developed an efficient implementation for MLEM on multicore architectures. In this work we seek for an efficient implementation of the MLEM algorithm on Xeon Phi (Knight’s landing) using hight-bandwidth memory (HBM). Verification is to be done by benchmarking against the Intel Math Kernel Library (MKL). A cluster system consisting of Xeon Phis is available at LRZ (CooLMUC3).
Contact: Tilman Küstner
Autoscaling performance evaluation on Cloudsuite applications
One of the directions of our research group's work is the evaluation of the autoscaling performance as it seems to be problematic to meet QoS requirements under rapidly changing load for example. We have developed a unique solution to evaluate the performance of different multilayered autoscaling solutions (e.g. autoscaling of AWS, Azure, Google cloud + Kubernetes horizontal scaling of pods). This solution, namely Autoscaling Performance Measurement Tool (APMT), was used until now only on the single application which is CPU-intensive, and we want to test the autoscaling for various application types using this tool on the benchmark Suite of EPFL that covers most of the application types.
In the scope of the bachelor work you will need to adapt the applications of the Suite for the tool, collect the data, and analyze it to detect autoscaling patterns and performance issues for different application types. The results of such work will help us understand the major drawbacks of autoscaling solutions in respect to particular application types. These results could be integrated in the paper, where you will be a co-author.
Contact: Vladimir Podolskiy