This session will present a memory scale-out solution that liberates SAP HANA, or similar memory demanding enterprise applications, from the classical limitation of underlying physical servers. The solution relies on key enabling technology developed within SAP. It allows applications or hypervisors to go beyond the boundaries of the underlying hardware, and effectively enables a fluid transformation from commodity sized physical nodes to very large virtual instances, in order to meet the rapidly growing demand of memory intensive applications.
Main point: Networking faster in comparison to other latencies. The immediate conclusion would be to scale-out.
Walk : sequential => each thread start reading from 256k / nb threads * threads id and ends when it reach the start of the following threadsBinary split : we split the memory in Nb threads regions . Each threads will then do a binary split walk within each regionRandom walk : each thread will read a page randomly chosen within the overall memory region ( no duplicate)We Are maxing out the 10GbE bandwith with IWARPWe suspect that we do not have enough core to saturate the QDR linkWe have almost no noticeable degradation when we have Threads > CoresSoftIwarp has a significant overhead ( CPU – latency- Memory use)