Cloud extensions, part 1th: Building compute nodes or small cluster applications and scaling with HPC

Use warehouse size (Warehouse-scale) calculations as needed

Processor cores with custom extensions and shared memory the external HPC architecture of the internet is rapidly being replaced by an on-demand cluster, these clusters take advantage of off-the-shelf general-purpose vector collaboration processors, converged Ethernet (each link-gbit/s or higher speed), and multicore headless (headless) servers. These new on-demand HPC resources are similar to the so-called warehouse sizing (Warehouse-scale computing), where each node is homogeneous and headless, focusing on total cost of ownership and overall power usage efficiency. However, HPC has the processing power needed to resolve beyond similar social networks, WEB searches, and other typical warehouse sizing solutions. This article will focus on how system builders and HPC application developers can extend your systems and applications most efficiently.

Migrating to High-performance Computing

Since 1994, TOP500 and Green500 supercomputers are not typically custom designed, but are designed and integrated using off-the-shelf headless servers, converged Ethernet (or InfiniBand clusters), and generic graphics processing unit (GP-GPU) coprocessor, which are not For graphics processing, but for single program, multiple data (SPMD) workloads. High-performance Computing (HPC) deviates from the direction of external custom processor and memory interconnect design, with trends towards leveraging existing equipment (warehouse size calculations) (based on controlling TCO, increasing power efficiency, and balancing operating expenses (OPEX) and capital expenditures for new and established HPC operations (CAPEX) Needs). This means that you can build your own small clusters in a similar way, using them as needed when you need HPC warehouse size resources.

The famous 3D ring interconnect used by computers such as Cray never completely disappeared (today, TOP500 1/3 uses a large parallel processor [MPP],2/3 High-performance machine uses a cluster architecture), but for efficiency and new OpEx metrics such as Green500 floating point The focus of Operation (flop)/watt) is driving HPC development and keeping the architecture focused on the cluster computing architecture. Also, many interesting applications today are data-driven (e.g., digital video analytics), so many systems need not only to use traditional sequential high-performance storage for HPC checkpoints (the saved state of long-running jobs), but also to randomly access structured (database) and unstructured (file) large datasets. Large data access is a common requirement for traditional warehouse sizing and current and emerging HPC workloads for cloud services. As a result, warehouse sizing is not HPC, but HPC applications can use the technology inspired by the data center to implement on-demand cloud HPC, provided it was designed from the outset.

Calculate power

The calculated power can be measured in a standard per watt performance form-for example, Flops/watt or per-watt input/output for calculation and I/O, respectively. Moreover, any computing device can be considered a factory that converts watts to computational results, and the overall measurement of excellent plant design is power efficiency (PUE), in short, the ratio of total energy consumption to computing equipment. Now, this value is not more than 1.2 is very good. One reason for the higher PUE is inefficient refrigeration, management overhead, and a lack of customized equipment compared to cloud data centers.

The focus of the scalable computing architecture is always changing, including:

Early attention to the use of a fast single processor, the storage program arithmetic logic of the unit CPU to the highest clock rate and instruction throughput:

John von Neumann, Alan Turing, Robert Noyce (Intel's founder), Ted Hoff (Intel Universal Processor advocate) and Gordon Moore, Consider the initial extension as a big challenge to extend the digital logic and processor clock as much as possible.

At least until 1984 (and possibly later), people generally think that "the processor makes the computer."

Cray Computer has designed vector computers (X-MP and Y-MP) and distributed memory multiprocessor, which are interconnected by a 6-to interconnected 3D loop for custom MPP machines. But this is a unique design in the field of Supercomputing.

IBM's early focus was on scalable mainframes and fast single processors, until 1999, when the multicore IBM Power Architecture onboard system was designed and a 3D ring-connected IBM Blue Gene Architecture was launched. The current TOP500 contains a number of Blue Gene systems, which are often topped in the TOP500 of LINPACK metrics.

From 1994 until recently, HPC evolved, evolving into custom MPP and usually off-the-shelf clusters, using custom interconnections (such as Blue Gene and Cray) and off-the-shelf converged Ethernet (10G, 40G) and InfiniBand:

TOP500 has been ruled by the cluster, which includes today's most top performance HPC solutions (2/3).

As shown in the Architecture TOP500 chart since 1994, clusters and MPP now dominate (relative to single instruction, multiple data [SIMD] vectors, fast single processors, Chenduo to [SMP] shared memory, and other less-defined architectures).

John Gage of Sun Microsystems (now Oracle) shows that "the network is a computer", which refers to distributed systems and the Internet, but similarly, the low latency network in the cluster is becoming the core of the expansion.

A coprocessor that is connected to a cluster node through memory-mapped I/O, including Gp-gpu and a hybrid field Programmable gate Array (FPGA) processor, is used to speed up specific compute workloads on each cluster node.

Warehouse sizing and cloud start up, and they focus on highly parallel applications (embarrassingly parallel applications) that MapReduce and HPC call:

TOP500 is measured using LINPACK and flop, so it is not concerned with operating costs (such as Flop/watt) or data access. Memory access is important, and storage access is not so important, except for job checkpoints (so you can restart a job if needed).

A number of data-driven query applications have emerged in the new century, including social networks, Internet search, global geographic information systems, and analysis associated with more than 10 internet users. This is not in the traditional sense of HPC, but large-scale warehouse computing operations.

Luiz Andrébarroso indicates that "data center is the computer" and this is the second time away from focusing on the processor. Data centers are highly focused on OpEx and CapEx, so they are better suited to flop/watt and data access-critical HPC. Google Data Center has a pue,pue of less than 1.2 is a metric that will consume total energy consumed by dividing the power used for calculation. (Most computing companies have 2.0 or higher PUE, so 1.2 is really low.) ）

See more highlights of this column: http://www.bianceng.cnhttp://www.bianceng.cn/Servers/cloud-computing/