This week at SAPPHIRE NOW Dell is announcing its scale-up solution for SAP HANA. When discussing the announcement at Dell and with others in the industry it became apparent that many people are confused about the difference between scale-out versus scale-up. So in the next few paragraphs I will discuss the differences based upon Advanced Computer Architecture and Parallel Processing, which is the seminal reference in advanced computer architecture.
The term “scaling up” means to use a more powerful single server to process the workload that fits within the server boundaries. For example, by moving to a PowerEdge R920 you are moving to a machine that has more cores on Intel processors. We believe many of the HANA workloads can be handled by a single scale-up server at 2TB of memory. Even with this, there may be a point where scaling up reaches boundaries and the workload grows beyond what a single server can process. That is the reason that we have scale-out.
Scale-out is a different model which utilizes multiple processors as a single entity so a business can scale beyond the computer capacity of a single server. The way we accomplish this at Dell is by using technology that was enabled for high performance computing. It is very important that we not only look at the workloads, but at the characteristics of the amount of data that HANA or any other platform might need. We do this by incrementally adding clusters as a workload increases. This is the concept of using relatively small clusters rather than trying to fit the workload in a single scale-up system. This is by far more economical and enables greater scale than a single scale-up system in the right circumstances. We are working on a scale-out solution that will be available very soon upon completion of testing with SAP. Dell will be the first to offer HANA customers this choice.
When a customer is trying to decide between scale-out and scale-up HANA deployment it is important for them to determine how large the HANA database should be and performance characteristics. It’s not either/or but it’s directly inversely proportional to the workload and the cost-effectiveness. The main reason that customers go to scale-out over scale-up is to meet the scaling performance requirements of their environment. Scale-out allows you to combine multiple machines into a virtual single machine with the larger memory pool than a scale-up environment would need. In a scale-up you achieve higher performance over scale-out but are limited to the limitations of a single processor. Scale-up and scale-out do not perform in a linear fashion because the operational significance of the architecture makes scale-out slightly more complex. Put simply, you will not see double the performance of a single machine by adding two together. In a hot environment you will not see double the performance when two machines are combined, but it’s a small price to pay for the architectural overhead.