ASTRON & IBM Center for Exascale Technology

Face-to-Face meeting

From March 25-27, members from the international Dome team came to Dwingeloo for another face-to-face at ASTRON. It turned out to be an excellent meeting which showed great progress and a truly collaborative atmosphere.



After an update on the current status of the SKA project and the SKA preliminary design reviews, each workstream reported on their own progress and planned next steps. We have summarized some key points from each workstream for you.

Workstream P1 (Algorithms & Machines) has created a working tool for design space exploration. It was reported that this tool can evaluate 350 design alternatives of a specific processor within 50 seconds.

P2 (Access Patterns) has shown the proof-of-concept for the access pattern prediction method working, has working tools for storage system sizing and is asked to move forward with incorporating SKA's objective of regional data centres.

P3 (Nano Photonics) showed first measurement results of an all-optical beamformer, while P4 (Microservers) reported on their proposed approach for the MeerKat telescope. Be it with some delay, we are also approaching delivery of a number of T4240 microservers to the Users Platform.

P5 (Accelerators) reported on an impressive set of implementations of the correlator in GPU/DSP systems and on the new GPU-based imager. Near-memory accelerators are moving to prototype their implementation on IBM Power-8 hardware.

P6 (Compressive Sampling) reported on novel algorithms and good initial results and performance for calibration and imaging. The latter uses an orthogonalization approach based on Gram-Schmidt, without the need for gridding or the FFT.

P7 (Real-time Communication) reported on a prototype RDMA protocol specifically tailored for carrying radio-astronomic data in a partially reliable fashion. RDMA stands for Remote Direct Memory Access. As mentioned earlier in this newsletter, first measurements hint towards a significant reduction of power consumption for RDMA data transfer compared to using the TCP/IP protocol suite.

Also, we see an area of particular interest developing: the dimensioning of regional science data centres. These centres are not covered by the current SKA designs, but the dimensioning of these, and the data transport costing, is an essential step towards Dutch and European science data centres. P2 will study the SKA data transport cost for the centres.