Michael Sevilla: Using Phoenix for a Scale-up vs. Scale-out Comparison

Monday, March 18, 2013

When data gets too large, we scale to bigger systems, either by (1) scaling out (adding nodes to a system) or (2) scaling up (adding resources to a single node).

Previous scalability studies enumerate the tradeoffs between scale-out and scale-up but the studies use narrow methodologies and out-dated hardware. In the past (early 2000s), scale-out architectures garnered extensive attention in both the research and the business communities for three reasons: (1) non-linear system/hardware scaling, (2) cost, and (3) interoperability issues. At present, we are seeing more and more challenges for scale-out architectures, resulting in workload specific distributed architectures, optimizations made at the application level, and incredible complexity/unpredictably at the system maintenance level. In light of this, we propose a new scale-out vs. scale-up comparison to re-examine the benefits and drawbacks inherent in each architecture.

We will apply current distributed systems benchmarks and workloads to a powerful single node, predict their behavior, and compare the actual performance to their distributed systems counterparts. It is our hypothesis that scale-up systems will degrade differently than scale-out systems and we hope to quantify and explain these behaviors. This talk discusses the limitations of previous studies, our long-term and short-term goals, and the progress we have made in (1) selecting applications, (2) porting applications, and (3) applying new methodologies to measure performance.

PDF icon msevilla_scale-out-vs-up.pdf1.08 MB