Green Compute – More Workloads Less Power Less Space
Green Networks – High density switches that deliver more bits per power consumed
Green Storage – Save Your Electricity Bills With Energy Optimized Storage
Green Compute – More Workloads Less Power Less Space
With current geo-political situations that cause high energy prices, electricity rates are now near record highs. There are 3 major clusters of equipment that make up a computing infrastructure: the network, the compute, and the storage. Each of these clusters of equipment consume vast amounts of power, and generate a correspondingly large amounts of heat.
In our previous newsletter, we discussed about green networks. In this newsletter, we look at compute, and how it can contribute towards green computing, and carbon neutrality.
High Performance Computing
In August of 2019, AMD introduced to the world the EPYC 7000 series of 64-core count CPUs, codenamed Rome. In comparison, Intel’s equivalent at that time was the Xeon’s series, with core count of 28 to 56 cores.
The availability of high-core count CPUs unleashed a wave of innovation in high-performance computing, and a shift in design thinking of compute infrastructures. Server manufacturers began to introduce high-density servers based on AMD’s Rome CPUs. For example, a 1U-sized server can support up to 2 CPUs, delivering 128 cores in a single server node. Other variants are 2U-sized, with 4 drawer nodes of servers, each with 64-cores, delivering up to 256-cores in a compact 2U space.
The availability of such high-performance, high-density servers meant that cloud providers and large enterprises can now build high-capacity compute in much less space, and consuming less power per core-count. When used with virtualization software, more workloads can be supported in a single physical instance of server.
What used to be racks after racks of servers could be compressed into less space, consuming less power for the equivalent in capacity.
This performance envelope in CPU is expected to be exceeded in the next lineup of CPUs, codenamed Genoa by AMD in 2023. We should expect to see even higher core-counts, and even denser compute.
Server Optimization
In order to drive more value out of servers, server optimization is implemented. The underlying cloud operating system, including the CPU virtualization, consumes a lot of CPU resources in handling tasks such as networking, storage, encryption and compression. Servers can be augmented with DPUs (Data Processing Units), a type of NIC cards with onboard co-processors that are very good at packet processing. With a DPU installed in a server, tasks such as networking (or storage, encryption/compression) can be offloaded from the CPU to the DPU. Offloading frees up the CPU to take on more user workloads, and speed up the overall performance of the server.
Another server optimization is the implementation of GPUs (Graphical Processing Units), to take over mathematically compute intensive tasks for workloads such as AI or machine learning. GPUs are very good at vector processing and parallel processing.
Lower Total-Cost-of-Ownership
If an enterprise has a large compute infrastructure that is currently being supported by racks upon racks of compute, there is an opportunity to compress the compute into less racks, and consuming less power for the same capacity. With cloud operating-system such as OpenStack, we can build hyperscale compute that scales linearly, delivering additional performance by adding on more compute nodes. Such design methodologies can result in savings on hardware (less nodes), software (less licensing), less space, and less power and cooling for the same capacity.
High-performance and high-density compute can save a lot of rack space. However, we have to be careful to design a balanced system that is within the limits of power and cooling in a single rack. It is also important to have fast compute connected to fast network and fast storage. Existing data-centers may not be designed or is able to accommodate all the high-performance compute infrastructures, when power or cooling is inadequate.
What’s Next
In our next newsletter, we will discuss about storage, and how it forms a complete green infrastructure with network and compute.