Improving network performances within the cloud – eduPERT and a real case
The ever-increasing complexity of applications and data is making the end-to-end performance more and more critical. Performance Enhancement and Response Teams (PERT) are at the forefront of technology and are a crucial resource.
eduPERT is a community open to all interested in performance that combines all the independent PERT teams. EduPERT is part of GEANT, the pan-European research and education network and it provides multi-disciplinary support to solve end-to-end performance issues including new technologies such as: high-speed network, big data, and new end point environments like cloud or data center.
Since September 2012 SWITCH, the Swiss NREN (National Research and Educational Network) and leader of eduPERT has been working on a Cloud project called Building Cloud Competence, or shortly BCC.
The goal of BCC is to learn as much as possible about Cloud and virtualisation by means of a pilot built in-house that can be leveraged to test different software solutions and architectures.
Twelve modern x86 servers are allocated in a rack, two of them acting as front ends. The external and internal connectivity is achieved through two top-of-rack (ToR) switches and bundles of dual 10GE links.
OpenStack and Puppet are used for automating the process of installing and starting up virtual machines (VMs) while the chosen virtualisation solution is KVM.
The execution of some iperf tests showed that the maximum throughput achieved among VMs was 110 Mb/s, a very low throughput considering the overall capacity of the infrastructure.
Investigating on the issue, we discovered that VMs in BCC are presented with an emulated "rtl8139" network interface that severely limits performance. The conclusion was that the use of a higher-performance emulated network device, preferably a para-virtualised adapter would improve considerably this result.
The migration to a para-virtualised approach from a fully virtualised one was already successfully used during another PERT case observed on the SWITCH DNS server, a VM itself.
Unfortunately the combination in the cloud of OpenStack nova-compute and KVM made the migration much harder and complex than we previously experienced with a single virtual machine.
In this presentation we would like to introduce an overview of eduPERT and a more detailed description of the BCC architecture together with the steps and the changes we applied in order to improve the overall network performance.
Alessandra Scicchitano's Biography Alessandra Scicchitano received a Dr.Ing. degree in Computer Engineering and Ph.D. degree in System and Networking in 2004 and 2007 respectively from UniCal, Italy. From 2005 to 2007, she was visiting researcher at Politecnico di Torino in a joint program for her PhD studies, working on scheduling algorithm for IQ switches. From 2007 to 2009 she held a PostDoc position at IBM Zurich Research Lab, working on the IEEE 802.1au standard and on algorithms for adaptive routing in HPC systems. Today she is part of the Peta Solutions team at SWITCH and her main focus is on virtualization and E2E performance. |