As some of you are already aware, we have suffered many unexpected delays in the setup of the
iris cluster - the fault lies to one of the solution provider of the computing RFP which failed to perform several early stage items in due time at all levels (hardware, network and to a minor measure software), blocking all dependent actions and delaying as well the other operators involved in the storage RFP and our own complementary configuration duties.
We had to take several drastic measures to accelerate the release of the cluster which came with a lot of additional efforts on our side. It was not an easy task however we are reaching now a more comfortable situation permitting the validation of the new platform despite a few missing network cables (pending delivery) required to ensure the Fat Tree. You will see below the result of several preliminary performance evaluation of the new cluster. First on the Infiniband network capacity, using MVAPICH OSU Micro-Benchmarks .
In terms of computing capacity, we have made a serie of runs over the HPL - High-Performance Linpack to predefine the set of input parameters foreseen to maximise the performance of the final execution over the full platform (once the network is completed). This stepwise evaluation is depicted below – so far we obtained 75 TFlops (using 95% of the resources) when we expect for the final run a measure closed to 85 TFlops.
Starting next week:
- the storage system will be also evaluated by DDN to ensure compliance with the performance expected from the RFP (i.e. at least 10GB/s RW)
- the missing IB cables will be delivered (ETA: May 19th) and installed to complete the FAT-Tree topology, leaving the opportunity to complete the performance evaluation of the computing nodes (using HPL and HPCG).
Then beta-tester will be welcome to complete the qualification of the cluster before its production release. If you are interested to take this role, kindly do not hesitate to contact us.
Finally, save the date of the next UL HPC School which will take place on June 12th and 13th in Belval. The program will be announced soon.