A look into the server consolidation in different contexts.
IBM Research just released a large-scale performance survey of several thousand data center servers over the past two years. The research shows that, on average, today’s servers’ biggest performance challenges are in-memory and disk utilization and not in processor usage, and that in about half the cases server consolidation will not make sense.
The study was a joint effort between Robert Bike and Lydia Chen of IBM’s research lab in Zurich and Evgenia Smirni of the College of William and Mary in Virginia. They studied a random selection of several thousands of servers hosted in different data centers, located in five continents, during the period of June 2009 to May 2011. Utilization values of CPU, memory, and disk in different time scales, (i.e., hourly, daily, weekly, and monthly) were all gathered. They characterized the consolidated server workloads in terms of resource demands and provided statistics that quantified demands across different resources. They then derived implications on the potential server consolidation strategies. Other factors, such as the effect on resource demands by geographical locations, time shifts and seasonality, were examined.
Some of the findings from the 13-page report included:
- Server consolidation isn’t always a good thing: As CPUs are averagely utilized at 15% to 20%, the theoretical upper bound of data center system improvement via workload consolidation is roughly six fold, given that the current utilizations are maintained and other resources are not the bottleneck. However, more than forty percent of the servers had peak CPU utilization over 50%, limiting any efficiency gains from consolidation.
- Memory and disk utilization are the key: The representative average workloads in current data centers are CPU (18%), memory (78%), disk (75%), and file system(45%).
- Bundled Resource Provisioning: Provisioning of CPU, memory, and file systems should be bundled together, and there is a good potential for resource saving in consolidating servers with heavily loaded memory and disk.
- Virtualization increases resource utilization: Virtual resources show a higher demand growth than supply growth, especially for CPU where virtualization technologies doubled CPU utilization. Virtual CPU utilization increased by 8.2% in two years (rising from 19.4% to 27.6%), whereas the physical CPU utilization only increased by 2% (rising 10.9% to 12.9%).
- Virtualization overhead hurt CPU performance – but does it matter? Difference between virtual and physical CPU utilization could be due to the virtualization overhead, but with CPU utilization PEAKING at just 27.6 percent for virtualized CPU, right now server managers will have bigger concerns than CPU usage. Both virtual and physical memories bear steady utilizations around 85% and 70%, respectively. On the contrary, virtual disks have about 6% lower utilization than the physical disks. For the file systems, both physical and virtual servers have similar utilizations around 43% with slightly higher utilization growth for virtual ones.
The full results of the research “Data Center in the Wild: A Large Performance Study” can be found on the IBM Research site here
