Electrical computers and digital processing systems: multicomput – Network-to-computer interfacing
Reexamination Certificate
1998-10-23
2002-01-08
Sheikh, Ayaz (Department: 2155)
Electrical computers and digital processing systems: multicomput
Network-to-computer interfacing
C709S226000, C709S223000, C709S224000, C709S238000, C370S351000, C710S003000
Reexamination Certificate
active
06338095
ABSTRACT:
BACKGROUND OF THE INVENTION
This invention relates to a data transfer method between plural element processors connected through an interconnecting network and a computer system suitable therefor.
A conventional parallel computer system often has structure that plural element processors each of which includes a local memory and an instruction processor are connected by an interconnecting network. In general, the parallel computer system of this type is called as a parallel computer system of a distributed memory type. Each element processor exchanges data stored in its local memory with other element processors by way of the interconnecting network and executes processing in parallel with each other.
In the parallel computer system of the distributed memory type, data transfer is generally achieved by using the programming model called message passing. In the message passing model, a sending (SEND) procedure and a receiving (RECEIVE) procedure are described explicitly in the parallel program (which is called as a user process hereinafter) executed in each element processor. The element processor of the sending side transfers a message which contains send data designated by a send procedure (SEND) when the procedure has been executed. The element processor of the receiving side receives the message when a receive procedure has been executed. The instruction processor in each element processor analyzes these communication procedures included in the user process under execution, and advances the processing by transferring data to the interconnecting network or receiving data therefrom. The element processor of the sending side designates the number of the destination element processor and transfers the message. Some specific data transfer methods are proposed by which each element processor actually processes the data send request and the data receive request from such a user process. From a practicable viewpoint, it is preferable that the extra processing (send overhead) executed from issuing of the send request by the user process of the sending side till the start of transferring of the user data and the extra processing (receive overhead) executed from issuing of the receive request by the user process of the receiving side till handing over of the data to the user process are small.
Recently the following data transfer method is used in many plural parallel computer systems, in order to reduce the send overhead and the receive overhead. The send/receive circuit of the element processor of the sending side directly read user data to be sent from a region for a user process provided in the local memory, and transfers the user data and generates a message which contains the user data. The send/receive circuit in the element processor of the receiving side directly writes the user data within the message into a region provided for a user process of the receiving side within the local memory. There are some specific methods for achieving such data transfer, but, in the present specification, those data transfer methods are generically called as a direct inter-memory data transfer method.
Specific examples of the communication method for executing the method are the PUT communication method for transferring send data to another element processor and GET communication method for receiving data from another element processor. For instance, refer to “Proceedings of Information Processing Society of Japan Parallel Processing Symposium JSPP '95” and PP. 233-240 (May, 1995) or refer to a manual “HI-UX/MPP Remote DMA User's Guide -C-” published by the present assignee for a parallel computer system SR2201 developed thereby. The PUT communication method is called as a direct remote memory access method in this manual.
To be specific, in the PUT communication method, the instruction processor in the element processor of the sending side notifies the send/receive circuit of an address (send data address) of the local memory of the sending side to read the send data, an address (receive data address) of the local memory of the receiving side to write the send data, and some other addresses. The send/receive circuit of the sending side reads the user data from the local memory based on the send data address, generates a message which contains the user data, the receive data address, and other addresses, and transfers the message. The send/receive circuit of the receiving side writes the user data within the message into the local memory in the element processor of the receiving side according to the receive data address in the message. Thus, in the PUT communication method, the send/receive circuit of the sending side directly accesses the local memory of the sending side according to the send data address and the send/receive circuit of the receiving side directly accesses the local memory of the receiving side according to the receive data address. Other communication methods which differ from such a PUT communication method but execute similar processing will be called below as a PUT communication method or a remote memory directly access method. The PUT communication method will be discussed below as a representative method of the direct inter-memory data transfer method, when the latter method is to be discussed.
In direct inter-memory data transfer method, the send/receive circuit which generates or receives a message may directly read the send data from a region for a user process in the local memory of each element processor or writes the received data into a region for a user process in the local memory. So, in the element processor of the sending side, the send data does not need to be copied from a region for a user process onto a buffer area (controlled by the Operating System), and in the element processor of the receiving side, the received data need not be copied from a buffer area (controlled by the Operating System) onto a region for a user process, either. Therefore, in this communication method, the send overhead and the receive overhead which derive from the copy can be reduced.
The communication library which executes the direct inter-memory data transfer method of such as the PUT communication or the GET communication and so on is developed by each parallel computer system maker or a research organization as a library specific to them. So, it is difficult to port the parallel programs programmed so as to use the library to other machines. On the other hand, a library (message passing library) with a standard interface specification to use the message passing model between user programs has come in use. It is possible to run such a parallel program programmed so as to use the library on different computers equipped with the library without change. The representative of that interface specification is MPI (Message Passing Interface). Universities in the United States of America and parallel computer system makers has organized the Message Passing Interface Forum (MPI Forum), and MPI is specification decided by the forum as a result of the research. It is expected that the library produced based on the specification will become a main current of the parallel program development support libraries in the future. Hereinafter, the library will be called as the MPI library.
In order to execute a data send request or a data receive request issued from a user process to the MPI library at high speed, it is effective to execute those request by using the direct inter-memory data transfer method like the PUT communication method and so on. Thus, parallel computer system makers and so on have developed MPI libraries to execute the send request or the receive request from the user process, by using the direct inter-memory data transfer library such as the PUT communication library. For instance, such a MPI library is used in the parallel computer system SR2201 developed by the assignee of the present application. The MPI library of this kind responds to the data send request or the data receive request from the user program, issues a suitable command to the PUT communicatio
Fujii Hiroaki
Yasuda Yoshiko
Lee Tammy
Sheikh Ayaz
LandOfFree
Data transfer method for reduced number of messages by... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Data transfer method for reduced number of messages by..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Data transfer method for reduced number of messages by... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2852129