Top 500 Computers Federated Distributed Systems Anda Iamnitchi.

Top 500 Computers

Federated Distributed Systems

Anda Iamnitchi

Objectives and Outline

• Question to answer in the end: differences between clusters and supercomputers?

• Top 500 computers:– List– How is it decided? (Linpack Benchmark)– Some historical data

• IBM Blue Gene

• Earth Simulator

System components BG/L

• Objective: 65,536 nodes • produced in 130-nm copper IBM CMOS 8SFG

technology. • Each node: a single application-specific integrated circuit

(ASIC) with two processors and nine double-data-rate (DDR) synchronous dynamic random access memory (SDRAM) chips.

• The SoC ASIC that powers the node incorporates all of the functionality needed by BG/L. It also contains 4 MB of extremely high-bandwidth embedded DRAM [15] that is of the order of 30 cycles from the registers on most L1/L2 cache misses.

• Nodes are physically small, allowing for very high packaging density.

System components BG/L (cont)

• Power is critical: densities achieved are a factor of 2 to 10 greater than those available with traditional high-frequency uniprocessors.

• System packaging: 512 processing nodes, each with a peak performance of 5.6 Gflops, on a doubled-sided board, or midplane, (20 in. by 25 in).

• Each node contains two processors, which makes it possible to vary the running mode.

• Each compute node has a small operating system that can handle basic I/O tasks and all functions necessary for high-performance code.

• For file systems, compiling, diagnostics, analysis, and service of BG/L, an external host computer (or computers) is required.

• The I/O nodes contain a software layer above the layer on the compute nodes to handle communication with the host.

• Partitioning the machine space in a manner that enables each user to have a dedicated set of nodes for their application, including dedicated network resources.

• This partitioning is utilized by a resource allocation system which optimizes the placement of user jobs on hardware partitions in a manner consistent with the hardware constraints.

Component Description No. per rack

Node card 16 compute cards, two I/O cards 32

Compute card

Two compute ASICs, DRAM 512

I/O card Two compute ASICs, DRAM, Gigabit Ethernet 2 to 64,selectable

Link card Six-port ASICs 8

Service card One to twenty clock fan-out, two Gigabit Ethernet to 22 Fast Ethernet fan-out, miscellaneous rack functions, 2.5/3.3-V persistent dc

2

Midplane 16 node cards 2

Clock fan-out card

One to ten clock fan-out with and without master oscillator 1

Fan unit Three fans, local control 20

Power system

ac/dc 1

Compute rack

With fans, ac/dc power 1

• Nodes are interconnected through five networks, the most significant of which is a 64 × 32 × 32 three-dimensional torus that has the highest aggregate bandwidth and handles the bulk of all communication. • There are virtually no asymmetries in this interconnect. This allows for a simple programming model because there are no edges in a torus configuration.

Principles:1. Simplicity for ease of development

and high reliability, e.g.:• One job per partition (hence, no

protection for communication)• One thread per processor (hence,

more deterministic computation and scalability)

• No virtual memory (hence, no page faults, thus more deterministic computation)

2. Performance3. Familiarity: programming

environments popular with scientists (MPI) for portability.

The Earth Simulator: Development Schedule

Development Goals

• Distributed Memory Parallel Computing System which 640 processor nodes interconnected by Single-Stage Crossbar Network

• Processor Node: 8 vector processors with shared memory

• Peak Performance: 40 Tflops• Total Main Memory: 10 TB• Inter-node data transfer rate: 12.3GB/s x2• Target for Sustained Performance: 5 Tflops using

Atmosphere general circulation model (AGCM). *The goal above was its thousand times' performance of

the AGCM in those times (1997).

System Configuration• The ES is a highly parallel vector supercomputer system of

the distributed-memory type, and consisted of 640 processor nodes (PNs) connected by 640x640 single-stage crossbar switches. Each PN is a system with a shared memory, consisting of 8 vector-type arithmetic processors (APs), a 16-GB main memory system (MS), a remote access control unit (RCU), and an I/O processor. The peak performance of each AP is 8Gflops. The ES as a whole thus consists of 5120 APs with 10 TB of main memory and the theoretical performance of 40Tflops

• Peak performance/AP: 8Gflops; Total number of APs: 5120• Peak performance/PN: 64 Gflops; Total number of PNs: 640• Shared memory/PN: 16G; Total peak performance: 40Tflops• Total main memory10TB

•Each PN is a system with a shared memory with:

• 8 vector-type arithmetic processors (APs): peak performance 8Gflops• 16-GB main memory system (MS) • remote access control unit (RCU)• an I/O processor

• Thus, 5120 APs with 10 TB of main memory and the theoretical performance of 40Tflops

Highly parallel vector supercomputer system with distributed memory:• 640 processor nodes (PNs)• connected by 640x640 single-stage crossbar switches.

The RCU is directly connected to the crossbar switches and controls inter-node data communications at 12.3GB/s bidirectional transfer rate for both sending and receiving data. Thus the total bandwidth of inter-node network is about 8TB/s.

The number of Cables which connects PN cabinet and IN cabinet is 640x130=83200, and the total extension is 2,400Km.

Figure 1: Super Cluster System of ES: A hierarchical management system is introduced to control the ES. Every 16 nodes are collected as a cluster system and therefore there are 40 sets of cluster in total. A set of cluster is called an "S-cluster" which is dedicated for interactive processing and small-scale batch jobs. A job within one node can be processed on the S-cluster. The other sets of cluster is called "L-cluster" which are for medium-scale and large-scale batch jobs. Parallel processing jobs on several nodes are executed on some sets of cluster. Each cluster has a cluster control station (CCS) which monitors the state of the nodes and controls electricity of the nodes belonged to the cluster. A super cluster control station (SCCS) plays an important role of integration and coordination of all the CCS operations.

Parallel File System

If a large parallel job running on 640 PNs reads from/writes to one disk installed in a PN, each PN accesses to the disk in sequence and performance degrades terribly. Although local I/O in which each PN reads from or writes to its own disk solves the problem, it is a very hard work to manage such a large number of partial files. Therefore, parallel I/O is greatly demanded in ES from the point of view of both performance and usability. The parallel file system (PFS) provides the parallel I/O features to ES (Figure 2). It enables handling multiple files, which are located on separate disks of multiple PNs, as logically one large file. Each process of a parallel program can read/write distributed data from/to the parallel file concurrently with one I/O statement to achieve high performance and usability of I/O.

Figure 2: Parallel File System (PFS): A parallel file, i.e. file on PFS is striped and stored cyclically in the specified blocking size into the disk of each PN. When a program accesses to the file, the File Access Library (FAL) sends a request for I/O via IN to the File Server on the node that owns the data to be accessed.

2 types of queues:• L: production run• S: single-node batch jobs (pre- and post-processing: creating initial data, processing results of a simulation and other processes

Programming Model in ES

The ES hardware has a 3-level hierarchy of parallelism: • vector processing in an AP, • parallel processing with shared memory in a PN, and • parallel processing among PNs via IN.

To bring out high performance of ES fully, one must develop parallel programs that make the most use of such parallelism.

• Yellow: theoretical peak; Pink: 30% sustained performance. The top three were honorees of the Gordon Bell Award 2002 (peek performance, language and special accomplishments). • Nov. 2003: The Gordon Bell Award for Peak Performance: Modeling of global seismic wave propagation on the Earth Simulator (5Tflops)• Nov. 2004: The Gordon Bell Award for Peak Performance:The simulation of a dynamo and geomagnetic fields by the Earth Simulator (15.2Tflops)

Cluster vs. Supercomputer?

• …