The present patent application is related to the benefit of the following co-pending China and U.S. patent applications: China patent application No. 200610029753.0, titled for “A Parallel Processing System with Self-Consistent Expandable Internal and External Networks”; China patent application No. 200610030472.7, titled for “A Self-Consistent Multi-rank Tensor Expansion Scheme and Multi-MPU Parallel Computing Systems”; China patent application No. 200710042397.0, titled for “A Mixed Torus and Hypercube Multi-rank Tensor Expansion Method”; China patent application No. 200610117704.2, titled for “Routing Strategies for Cellular Networks in MPU Architectures”. The whole contents and disclosure of the abovementioned related patents are expressly incorporated by reference herein as if fully set forth herein.
The present invention relates generally to the field of supercomputing systems and multiprocessor architectures, and more particularly to an ultra-scalable supercomputer based on MPU (Master Processing Unit) architecture.
With the rapid increase of a single processor processing ability, to construct a high-available, high-density and ultra-scalable supercomputer faces the embarrassing bottleneck of the communication subsystem design, whose developments and innovations dramatically lags behind the computation power. To better manage the consistent balance between the communication and computation in a large-scale supercomputer while to preserve the programmability and application portability, more and more advanced interconnect topologies are put into practice such as IBM BlueGene and Cray RedStorm employing 3-D torus and Columbia QCDOC employing truncated 6-D torus. Those novel multi-dimensional mesh-based architectures with the internally developed design of the advanced interconnect intellectual properties allow systems to comprise higher processor counts and to achieve better overall performance on applications compared to the conventional systems that rely on the commodity interconnect fabrics including the external fat-tree Infiniband or Myrinet federated switch. Additionally, as observed on systems with Myrinet and Infiniband interconnects, intermittent errors on a single link may have a serious impact on the performance of the massively parallel applications by slowing down the communication due to the time required for connection recovery and data retransmission. Alternately, mesh-based network designs overcome this issue by implementing a large portion of data links as traces on the system backplanes and by aggregating several alternative connections into a single bundle attached to a single socket, thus reducing impact caused by the number of possible mechanical faults by an order of magnitude.
Further, due to the absence of the more centralized external federated switch formed by a large number of individual switches, a mesh-based supercomputer is more flexible to be expanded without losing scalability. However, a similar-scale cluster design with several thousand nodes necessarily has to contain much more external cables connecting the nodes to tens of switches that must be located tens of meters or more away from each other. Such scenario is also a serious maintenance disaster, in addition to much higher operating expenses. The fault tolerance ability of a cluster system important for achieving the sustained and reliable computing resources for large-scale realistic applications is always guaranteed at the sacrifice of more redundant switches available, thus increasing the size of external federated switches again. However, the novel mesh-based supercomputers, such as IBM BlueGene, use smart adaptive routing strategies to isolate the fault area while routing messages with existing alternative data links around fault nodes flexibly without demanding much more hardware compensation to retain the communication balance by dynamic fault-tolerant routing algorithms.
The trends of the high-performance computing (HPC) industry strongly indicate the value of the present invention in providing an ultra-scalable supercomputer based on MPU architecture. The MPU architecture is a novel multi-dimensional network topology enabling an ultra-scalable and highly coupled interconnection while, by adding Axon nodes, providing facilitating long-range and collective communications and also for connections to external networks such as the management and external file networks. Preferably, considering both high compatibility with different processor platforms and flexible adaptability with varied application acceleration units while better and stably implementing our own IP cores, we choose the reconfigurable chip as the implementation platform of the switching logic for a processing node or an Axon node.
The object of the present invention is to provide an ultra-scalable supercomputer based on MPU architecture to achieve the high-performance and sustained computing resources at the scale of TFLOPS, PFLOPS and beyond at cost, power dissipation, and footprint advantages.
Another objective of the present invention is to provide a well-balanced supercomputer between the computation and communication abilities by the novel multi-dimensional interconnect topology as described in herein incorporated, pending China patent application No. 200610029753.0 entitled “A Parallel Processing System with Self-Consistent Expandable Internal and External Networks”. The said novel interconnect topology is used to achieve a network system with low network diameter and high bandwidth for the large-scale expansion.
A further object of the present invention is to provide a flexibly expandable and easily compatible supercomputer using the FPGA-based Inter-process communication Network (FIN) in achieving the ability of leveraging existing robust processor platforms for the computational functionality while focusing on the crucial internally developed interconnect topologies, in order to permit a high degree of seamless improvements with the continuing increase of the user-specific and commodity processors powering the system and to well accommodate compatibility to custom application acceleration units such as vector processing engines. Moreover, the FPGA chip for the router device can be either other reconfigurable chips or application-specific integrated circuits (ASIC) for the implementation of the system connectivity and further, a switching chip and processing cores with other supportive units such as multi-level caches and high-speed connectors can be integrated into one chip to implement a System-on-a-Chip (SoC) design.
The FIN design philosophy further enables designers, developers and end-users to adapt and upgrade the routing functions easily and flexibly. For example, as for a specific custom application, we can implement particular communication patterns in the reconfigurable chip to better balance and accelerate the program running and further, the reconfigurable chip can take over some computational loads off the local CPU subsystem without affecting other functional blocks' configurations and designs. Additionally, in a parallel computer with MPI, we can also implement and optimize basic communication routines and collective communication schemes into the routing chip to reduce the system overhead by better exploiting the intrinsic nature of the MPU-based interconnect architectures.
A further objective of the present invention is to build a multi-MPU supercomputer with a single high-density and cost-efficient processing cell as a supernode. A processing cell comprising several processing nodes and associated Axon nodes becomes a basic construction brick. Through the tensor expansion scheme described in herein incorporated, pending China patent application No. 200610030472.7 entitled “A Self-Consistent Multi-rank Tensor Expansion Scheme and Multi-MPU Parallel Computing Systems”, a ultra-scalable supercomputer is expanded while preserving a short network diameter, a high bisection bandwidth and sufficient alternative data links between processors for high bandwidth and strong fault tolerance.
The preferred hardware implementation of the present invention, a supercomputer system is based on the blade chassis solution considering the footprint and maintenances. The blade chassis solution enables the high-density multi-processor computer system at low cost, low power dissipation and high availability.
A preferred supercomputer at the present invention is to incorporate multiple interconnection networks. A first network implements a point-to-point MPU-based interconnect topology while supporting such collective operations as broadcasting and All-gather communication patterns. In said MPU-based interconnect topology, each of the processing nodes situates at the center of a multi-dimensional cube made up of its neighboring processing nodes and particularly, a boundary processing node situates at the center of a virtual multi-dimensional cube made up of its realistic and imaginary neighboring processing nodes due to the cyclic property of topology while the entire interconnect topology remains consistent and unified.
A second network implements an expansion network by connecting all of the Axon nodes for performing the long-range and collective operations such as global notification and barrier operations. In a multi-dimensional MPU topology, an Axon node connects a subset of processing nodes connected as an embedding of two multi-dimensional cubes of equal size in a preferred embodiment. Herein, these processing nodes are called as the Axon node's attached processing nodes and the Axon node is called as the upstream Axon node of these processing nodes. Next, all of Axon nodes according to their logical positions connect to form another small-scale MPU interconnect topology or a mesh-based topology instead. Additionally, an Axon node also provides its attached processing nodes with connections to external networks such as the management network and the storage file system.
A third network is the Ethernet-based management network for performing the remote monitor and administration operations to the entire system. A forth network is the high-speed external storage network. A processing node can get access to the external file system with the help of its upstream Axon node through the high-speed connections.
The interconnection network topology for the ultra-scalable supercomputer of the present inversion is based on the multi-dimensional MPU architecture connecting a plurality of processing nodes and Axon nodes by multiple independent interconnection networks. Herein, MPU, Master Processing Unit, is termed as a basic processing cell comprising a subset of processing nodes with their upstream Axon nodes and wherein the subset of processing nodes logically connected as an embedding of two virtual multi-dimensional cubes of equal size. The detailed interconnect topology definitions are described in China patent application No. 200610029753.0, titled for “A Parallel Processing System with Self-Consistent Expandable Internal and External Networks”.
An exemplary embodiment of 2-D MPU architecture, the simplest MPU topology, is shown in
Considering the contemporary technology and techniques, a preferred embodiment of 3-D MPU architecture is described in detail hereinafter. High-dimensional MPU architectures can be designed and implemented in a similar way with the technology progress in the manufacturing industries. Moreover, in the following descriptions, we will make use of a reconfigurable Field Programming Gate Array (FPGA) technology as the implementation platform of the switching logic for our design approach. Therefore, a FPGA-based IPC Network is termed as FIN responsible for all inter-process communications between processors in the entire system.
However, it is not only constrained to the FPGA and other programmable chips convenient for performing switching logics can be also utilized. Moreover, an Application-Specific Integrated Circuit (ASIC) chip can be developed and further a switching functional block can be embedded into a central processing subsystem to implement a SoC design, should volume reaches to a critical point.
In the preferred embodiment of 3-D MPU architecture, each processing node logically situates at the center of a three-dimensional cube made up of its eight neighboring processing nodes. For example in
As described in China patent application No. 200610030472.7, titled for “A Self-Consistent Multi-rank Tensor Expansion Scheme and Multi-MPU Parallel Computing Systems”, an Axon node is configured by connecting a subset of processing nodes for providing external interfaces and an expansion network is further built up by connecting all of the Axon nodes still as MPU architecture. Those processing nodes are named attached processing nodes to their Axon node and said Axon node is named the upstream Axon node to its attached processing nodes. Therefore, a second interconnection network can be built by connecting all of Axon node as MPU architecture mainly for the long-range and some collective communications such as global operations.
Additionally, a third interconnection network, defined as the management network for remote monitor and system administration, is also implemented by connecting Axon nodes to the external network such as an Ethernet switch. Herein, each processing node owns an Ethernet connection to its upstream Axon node's Ethernet switch for communication with the external management subsystem.
Additionally, a forth interconnection network, defined as the storage network for access to the external file system, is further implemented by connecting Axon nodes through high-speed channels to the external file subsystem such as Infiniband Storage Area Network (SAN). Herein, an Axon node owns expansion slots supporting Fiber or Infiniband channels to access and external file server switches and then with the upstream Axon node, the processing node can communicate with the external file subsystem.
Processing Node Overview:
A processing node provides the computation and communication ability comprising a central processing subsystem and a router device.
Functional Blocks of Processing Node FPGA (PNF):
Processing Node FPGA, PNF as shown in
The clock and reset module implements the functions to generate the reset and clock signals for the other modules.
The management module implements the array of system registers, handles read and write requests to the registers coming from the sRIO/PCI-E IP adaptation module, makes the contents of the registers available for the other modules that make use of their contents and collect status information from the other modules for storage in the registers, detects activity on data links and coordinates with the local processor subsystem. LEDs show status and events of interest for monitoring the state of the FPGA chip. The system registers are accessible from the local processor subsystem through maintenance requests as the module is connected to the maintenance port of the sRIO/PCI-E IP adaptation module. The array of registers includes parameters required by the switch fabric, status of Aurora links, caption of events such as temporary link errors, etc. Also, registers are used to keep track of the number of packets received/transmitted on each link, buffer fill-levels, error flags etc. The management module also generates signals for driving LEDs mounted on the PCB.
The sRIO/PCI-E IP adaptation implements the high-speed interface towards the local processor subsystem. The module separates maintenance traffic and other types of communication to different blocks; maintenance requests will result in read or write accesses to the management module and all other types of communication will be forwarded to the FIN protocol adaptation module.
The FIN protocol adaptation module wraps the original packets provided by the sRIO/PCI-E IP adaptation module in a proprietary packet format suitable for the FIN switch architecture. The proprietary packet overhead is removed when the packets are sent to the sRIO/PCI-E adaptation module for delivery to the connected local processor subsystem. The module also checks packets for errors and reports such to the management module.
The switch fabric module handles all packet switching within FIN and requires several parameters to be operational and reports errors back to the management module.
The Aurora IP adaptation module implements the high-speed Aurora link-layer point-to-point serial links that connect PNF with other processing nodes and also connect PNF with its upstream ANF. This module reports link status including errors to the management module as well. This module also includes the link-level flow control mechanism and the link-level retransmission mechanism to provide a stable link-level high-speed transmission platform for the other modules.
Block Diagram of a Processing Node:
A processing node comprises a central processor subsystem, a CPLD device, an Ethernet PHY, a FPGA-based router device and other supportive devices such as a boot flash, Power VRM, memories slots and a temperature sensor as shown in shown in
As for the block diagram of an Axon node, a preferred Axon node mainly comprises a central processing subsystem, CPLD, an Ethernet PHY, an Ethernet switch as marked in red area, PCI-Express interfaces as marked in red area, a FPGA-based router device for multiple communication networks, a power VRM, several memory slots, a boot flash and a temperature sensor. The same devices as in the processing node hold the similar functions in the Axon node. Said Ethernet switch connects the Ethernet connections from all of its attached processing nodes for the management network. One or more PCI-Express slots are provided for the external networks. Optionally, one or more PCI-Express expansion cards (Fiber channel or Infiniband) can be inserted into the PCI-Express slots. A FPGA chip is equipped on the Axon node board for the communications between attached processing nodes and external networks including the management network, the storage network and the expansion network for the long-range and collective operations.
Axon Node:
An Axon node is an expansion node to strengthen the communication and management ability comprising both internal interfaces to attached processing nodes and expansion interfaces to other Axon nodes' router chips and external supportive networks. As described in the patent entitled “A Self-Consistent Multi-rank Tensor Expansion Scheme and Multi-MPU Parallel Computing Systems”, an Axon node directly connected to attached processing nodes can be seen as the first-level or one-rank Axon node in the multi-rank tensor expansion scheme while providing interfaces to other one-rank Axon nodes and consequently the expansion network by connecting all of said one-rank Axon nodes as a MPU interconnect topology is defined as the one-rank expansion network to facilitate the long-range communications. Next, a two-rank Axon node further directly connects to a subset of one-rank Axon nodes at the similar manner for providing interfaces to other two-rank Axon nodes. The said subset of one-rank Axon nodes logically forms an embedding of two multi-dimensional cubes of equal size. Also, a two-rank Axon node can share most of functional blocks with a one-rank Axon node at design and implementation. Therefore, another expansion network by connecting all of said two-rank Axon nodes as another MPU interconnect topology is defined as the two-rank expansion network to further reduce the network diameter. Iteratively, a multi-rank tensor expansion network subsystem can be constructed to improve the performance of the global operations. Since the multi-rank expansion scheme is the same as that of the one-rank expansion network, the descriptions hereinafter concentrate on the implementation of the one-rank expansion network or the expansion network for short.
Functional Blocks of Axon Node FPGA (ANF):
Axon Node FPGA, ANF as shown in
The differences between ANF and PNF are that ANF owns external expansion interfaces to multiple external networks. In a multi-rank tensor expansion scheme, a one-rank ANF has also high-speed interfaces with its upstream two-rank ANF.
Block Diagram of an Axon Node:
An Axon node comprises a central processor subsystem, a CPLD device, an Ethernet PHY, an Ethernet switch, a FPGA-based router device, PCI-E expansion interfaces and other supporting devices such as a boot flash, Power VRM, memory slots and a temperature sensor as shown in
Layout of a Generic Blade Node:
Considering the system reliability and footprint, a blade node layout hosing multiple generic nodes is illustrated in
Layout of a System Chassis:
A 7U system chassis, as shown in
Layout of a System Rack:
Six 7U system chassis are populated in the standard 48U system rack with hot-swappable fan blowers on the bottom and on the top, providing the airflow for the bottom three chassis and top three chassis respectively as shown in
A Multi-MPU Supercomputer:
A multi-MPU supercomputer is a multiprocessor computing architecture comprising a plurality of MPU-based supernodes interconnected by a multi-rank tensor expansion communication subsystem, thus featuring a self-consistent multi-rank MPU-topology scheme. The hardware embodiment of the entire supercomputer consists of a blade node directly inserted into a system chassis while multiple chassis are mounted into a standard rack with cooling and power suppliers, for high reliability and high availability with small footprint, at low cost, and low heat dissipation. Meanwhile, the hardware embodiment of a reconfigurable chip for performing communication guarantees high portability to existing parallel applications and the easy-to-update ability to keep seamlessly consistent with both the technology progress of commodity processor families and the algorithm improvement of customer applications.
Furthermore, a hybrid supercomputer can also adopt a conventional multi-dimensional torus or hypercube interconnect topology for performing the trunk communication system while increase another expansion network by adding Axon nodes through the tensor expansion scheme mainly for facilitating the long-rang operations, as described in China patent application No. 200710042397.0, titled for “A Mixed Torus and Hypercube Multi-rank Tensor Expansion Method”. Herein, a MPU is a processing cell comprising a subset of processing nodes connected on the conventional multi-dimensional torus network and an Axon node connecting to those processing nodes while providing expansion interfaces to external networks. An expansion network is built up by connecting all of said Axon nodes on the multi-dimensional MPU network or the conventional multi-dimensional torus network. Moreover, these said Axon nodes can be termed as one-rank Axon nodes and another set of Axon nodes serving as two-rank Axon nodes can directly connect to one-rank Axon nodes so, similarly, a two-rank expansion network is constructed by connecting all of two-rank Axon nodes. Iteratively, a multi-rank tensor expansion communication subsystem is constructed with the aid of Axon nodes. Meanwhile, Axon nodes also provide the external connections as mentioned above. The heterogeneous system combining a torus topology and the multi-rank tensor expansion scheme is the most straightforward embodiment while sharing most of the above-mentioned exemplary embodiment. However, an ultrascalable supercomputer integrating high-dimensional MPU topology and multi-rank tensor expansion scheme is comparatively preferred.
The routing strategies in MPU architecture can use the switching functions as described in China patent application No. 200610117704.2, titled for “Routing Strategies for Cellular Networks in MPU Architectures”, while enables a deadlock-free adaptive routing pattern in MPU architecture and better support the exploration of intrinsic merits of the MPU architecture.
Number | Date | Country | Kind |
---|---|---|---|
2007 1 0044230 | Jul 2007 | CN | national |
Number | Name | Date | Kind |
---|---|---|---|
7844796 | Vorbach et al. | Nov 2010 | B2 |
20040103218 | Blumrich et al. | May 2004 | A1 |
20070276959 | Yancey et al. | Nov 2007 | A1 |
20080003948 | Mitran | Jan 2008 | A1 |
20090024829 | Deng et al. | Jan 2009 | A1 |
Number | Date | Country | |
---|---|---|---|
20090094436 A1 | Apr 2009 | US |