Low-Power Hyperscale Computing
November 27, 2012, Server Design Summit, Santa Clara, CA—Partha Ranganathan from HP presented
The trends and opportunities for this next-generation data center are becoming well defined. The performance will increase from 100 to 1,000 times today’s levels to meet the growing demands for computation. Power will become one of the factors in the new metrics, to become MIPS per total power. One issue is that the cloud is one of the five largest electricity consumers in the world. The costs of power are split in half for use and cooling.
Data centers will need to have sustainable performance. Systems are not very scalable when there are 82 million servers in the cloud. Any changes in deployment run into a management wall, and the reliability of the technology, the integration, and the security comes into question.
Costs of acquisition and operation have to drop to one tenth of their current levels for the systems to expand. The growth markets in the emerging economies need high-volume, low cost systems that can operate on an income of a penny per person per month. These drives will push the industry to “green” clouds that have much higher performance at a small fraction of the costs of a typical data center.
On top of the needs for improved efficiency and lower costs, we have to consider the impact of highly improbable events. Emerging technologies can completely change the picture, and new technologies emerge all the time. The growing gap between the number of transistors possible in a single chip and the performance in compute, communications, and storage continues to grow. The move to multi-core processors is the first step in a fundamental change in architectures for compute systems. The advances in SoC and GPU designs is also leading the way to more dark silicon, sections of the chip that are powered down when not in use.
Storage and memory have been the system bottleneck for a while. The transfer rates and latencies are forcing systems to move more data to the active memory, but the memory capacity ratio is not growing as fast as it used to grow. One alternative is to move to 3-D stacks of memory and logic to increase bandwidth, reduce power, and increase the level of integration at a given cost.
Other memory technologies, especially in the non-volatile space, can be a good way to improve access times without needing to go to another device. 3-D non-volatile memories, like memristers, can be established as byte-addressable, persistent, low-energy memories that fit between the cache, SSDs and HDDs in the performance and capacity spectrum.
Communications are becoming pin count and bandwidth limited. As a result, data communications are moving to photonics. The ability to multiplex wavelengths in a fiber increases the available bandwidth of a single channel. The challenges are to get the costs down to increase the volumes of channels available. as a result of the changes noted, future systems will use ions in memristers for storage, photons for communications, and electrons for computation.
The assumptions for the changes are that the emerging data-centric systems will have data volumes growing much faster than compute, and compute will move to the data. Big, rich data will have much higher complexity and dynamism. Data will have static, synthetic, and personalized aspects. Actionable line feeds, prescriptive analytics, and ad-funded functions will be a part of the data assembly.
The increased data functionality will require more complex I/O moves and cross correlation to the larger data repositories. These data will be fast, total, and require deep analytics. The changing technologies and workloads present an opportunity for the industry. The implications of these changes will be to affect primary assumptions about systems. Beyond the box will be distributed systems, beyond the platform will be a new mixture of hardware, software, and apps. Beyond performance will be total cost of ownership.
Development of these new systems will require cross-functional teams as compute systems become ensembles of machines that are configured on a as-needed, when-needed basis. As a result, the industry will see many more micro –blades using low-power mobile SoC processors. Memories will dis-aggregate to a common pool of shared memory.
Power management will drive towards the average loads, with details of peak thermal, network, and electrical demands. The issues are in scaling the power to the number of servers in the systems and the interactions of the machines across the peaks. As a result, data center operators will need to have ensemble optimizers to manage the server arrays.
