Present: Alastair, David Acreman, Richard Bower, Benedict, Rogers, Peter Draper, Tobias Weinzierl, Tom De Vuyst, Jonathan Frawley, Stephen Longshaw, George Fourtakas, Abouzied Nasar, Matthieu Schaller, Scott Kay, Edo Altamura. Action: email around working group meetings that are known about. Action: Content for outreach website Action: Progress company NDAs Action: AGB to chase company. Completed actions: Action: Richard Set up coordination between working groups - paused for now Action: Richard Swift training day in February - workshop in Feb/Marc Action: Ali contact Olly Perks at ARM - Done Action: Ali setup NDA with networking company - Done - General updates Tobias - involved with task-based Met Office call. And Exeter. Discussion about future exascale ambition for particle hydrodynamics workshop. Arup - coastal erosion. Airbus - too busy. EDF - Damian Vielo base person to contact. But likely to be too busy. Action - Richard/Stephen - send out email to advertise (event bright and sph-sig), and mention that if others want to give a short talk, to let us know. Date: 25th Feb. Shorten talks to 20 mins, make it 2 hours long, time for chat at end. Make it a series, and have a bit longer to set up the next one. Stephen to give a talk about SLOAD project next time. Next one to be 2 months time. - Excalibur funding call Tobias bid for the tasking call. Dave - talked about link between parallel-in-time and tasking calls. Link is the test cases. Having a think about how might to this mathematically, seems to fit with call ask, development in one area and how this fits in other areas. EPSRC call not yet released. This cross cutting themes will be released on 16th Feb (approx). Won't be overlap with Met Office. Refined to 3 page summary. - Contracts (Richard) Ongoing. - Hydro weak-scaling (Edo) Weak scaling tests done on COSMA7. Much larger this time. 100-356 nodes. Two variants - Intel MPI and OpenMPI. Scaling looks good from 2000 cores upwards. But we don't know below this. IntelMPI 2018 successful, but OpenMPI not so successful - some jobs hung. With 256^3 per tile, also looks good, scales poorly initially, but then improves. 128^3 uses 0.8% memory. 256^3 uses 5% memory. One additional thing tried but failed was to increase particle number further. 100Bn but didn't have enough time. All of cosma7 should allow 500Bn particles. - RDMA updates (Peter) Last time talking about RDMA changes - trying to work out what was going on - turns out to be explainable - problem with the Physics, rather than the implementation. Single core memory bandwidth limit is being reached. Separated out send and receive tasks, and use 8x more tasks, which gives a ~3x speed up. This mainly helps steps without much physics going on. Good side is that a similar change to the existing swift code also sees similar gains. SWIFT Richard plot - many time steps, amount of work v time taken per step. Fastest step is approx 50ms when there is very little to do using the RDMA code, and 100ms when using MPI. Less scatter in the RDMA version too. Need to decide how to do this properly now, as existing method is a bit of a hack. - GPU update (Abouzied) OO code to optimise cell interaction on GPU. Each object has N particles, nested objects. Each cell has heighbour objects which also contian the particles. Some memory duplication, but should run more smoothly. Getting about 100x speed up (without coms) or 35x with comms. But looking to develop async version, which should give the 100x. On a low power Quadro. - Individual workpackage update [all] - Training (Tobias?) - Outreach (?) Website needs setting up. - Others/AOB?