Linley Data Center Conference – Changes in the Cloud
February 7, 2012, Santa Clara, CA – At the Linley Data Center Conference both systems and chip suppliers were disucussing the new directions in data centers and cloud service configurations. As the mix of content that is on the internet changes, so does the optimization of the hardware to deliver that content.
In the general overview of the market, the analysts presented that the next stage in evolution of the network is the cloud, in the form public, private and hybrid configurations. This evolution come from the chain starting at main frames moving to client/server and now cloud. These cloud solutions have different functions, and hence different architectures. Public clouds are for web infrastructure and IT infrastructure. Private clouds are for CRM, ERP, Data warehousing. There is a split based on function for those doing computing in the cloud, and the public/private split is based on the where the results are used and viewed. This step tends not to embrace virtualization as computation is a machine directed function.
The majority of the discussion was on virtualization, the creation of hypervisors for the processors and offload of the main CPU with co-processors for specific tasks. The energy efficiceny of the data centers and the throughput are the keys. For energy efficiency using virtual machines spun out out software are the lowest energy path to more available resources. These include Vmware’s vSphere, RedHat’s KVM and Citrix’ XenServer. The Vms are used in conjunction with standard chip pwoer saving such as DVFS, and turning on/off blocks and cores that are unused based on data flow.
The connectivity side showed and discussed the migration from 1Gbps to 10Gbps and the rack confections are moving from 10G to 40G connections. The 40G configuration os 4 lanes of 10G based on the minimization of cabling costs. This same driver is pushing 100G from 10lanes of 10G to 4 lanes of 25G. The reduced cost for the number of lines, and the associated overhead of the 10 line setup, has significant impacts on the pricing of the connection path.
There was a discussion on L2 and L3 networking issues and offload hardware which can perform network packet processing (such as deep packet inspection, and encryption/decryption) without impacting the main CPU. This went in the direction of OpenFlow and capabilities of OpenFlow controllers.
The final topic of the overview discussion was switches. These bocks are increasing in desnsity in addition to speed. The typical configuration is 48 x 10GbE, while at the upper end Marvell has a 96 x 10GbE device. For new products – up to 32 ports of 40GbE at ta power of 1W per port oare starting to be seen. These switches are now starting to add the FCoE support directly in the chips and many of the OEMs are starting to developa core fabric internally to their products.
