Electrical computers and digital processing systems: multicomput – Computer-to-computer data routing
Reexamination Certificate
2001-08-01
2004-11-30
Lane, Jack (Department: 2188)
Electrical computers and digital processing systems: multicomput
Computer-to-computer data routing
C709S232000, C709S213000, C709S216000
Reexamination Certificate
active
06826622
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention concerns to methods of communicating data between computers in a computer system having a plurality of computers or data processing equipment connected through a communication network. More particularly, the invention consists of a method of sending/receiving data between memories of computers on a network in a which the hardware has the capability of transferring data between the memories of these computers.
2. Description of the Related Art
The TCP/IP protocol is used in the overwhelming majority of communications between computers, in particular in the communications in the Internet or in intranets. Since TCP/IP processing is not executed by the application, but is executed by the operating system, in order that the application perform communication using TCP/IP it uses an API (Application Programming Interface: the set of functions which an application calls in order to use a certain function of a computer or an operating system) called “Sockets API” (refer to the book by W. Richard Stevens, “UNIX Network Programming”, Prentice Hill, U.S.A., 1990, ISBN 0-13-949876-1).
An example of the software structure of a host which performs communication using the TCP/IP protocol is shown in FIG.
1
. The host
10
performs communication using the network
18
. The kernel
120
of the operating system of the host
10
executes protocol processing
121
of TCP/IP and controls the communication hardware
11
in order to perform communication. The program
101
of the application
100
uses the Sockets API
90
to call the library
110
. The library executes the system call
111
and calls the kernel
120
. The kernel
120
sends and receives data
102
of the application
100
through the socket buffer
122
.
Since protocol processing
121
in TCP/IP communication involves a large amount of processing, and the system call
111
and the copy between the data
102
and the socket buffer
122
result in overhead, these processings limit the communication performance in some cases. For this reason, computer systems requiring high communication performance, such as supercomputers or workstation clusters, employ networks which can transfer data between applications without performing protocol processing, system calls and data copies and also bypassing the kernel. In the present specification, henceforth, this communication method will be referred to as “high-speed communication” for short, when applicable. As an example of high-speed communication, there is the VIA (refer to the specification by Compaq Computer Corp., Intel Corp., Microsoft Corp., “Virtual Interface Architecture Specification, Draft Revision 1.0”, Dec. 4, 1997, http://www.Viarch.org). Since the functionality of high-speed communication is different from that of TCP/IP, their respective APIs are also different.
An example of the software structure of a host employing high-speed communication is shown in FIG.
2
. The program
104
of the application
103
calls the high-speed communication library
130
by using the high-speed communication API
91
to send and receive data
105
. By executing the communication processing
131
of the high-speed communication library
130
, the high-speed communication hardware
12
is activated bypassing the kernel
120
to send and receive the data
105
through the high-speed communication network
19
. When sending and receiving data by high-speed communication, two processings are required: the processing of inspecting whether or not the application
103
has the permission to access the data
105
which it wants to send or receive, and the processing to convert the virtual addresses which were specified by the application
103
into the physical addresses which are used by the high-speed communication hardware
12
. For this reason the application
103
, before sending and receiving data, calls the high-speed communication library
130
to register the data
105
to be sent and received (the registered data is shown in the form of a rectangle having rounded corners). The kernel performs the registration processing
123
in response to the call
132
of the high-speed communication library. As a result, it is possible to verify if the application
103
has access permission and, when it has the address conversion is performed and its result is registered in the memory registration table
13
. The high-speed communication hardware
12
performs both the verification of the access permissions and the address conversion by using this memory registration table
13
.
Since the high-speed communication API
91
is different from the Sockets API
90
, in order that an application
100
employing the Sockets API
90
may use high-speed communication, this application
100
must be rewritten to use the high-speed communication API
91
. Since this rewriting is difficult to do, many applications will remain unchanged, still using the Sockets API, thus they won't be able to take advantage of the high performance of high-speed communication. In order to solve this problem, a communication method called “Fast Sockets”, shown in
FIG. 3
, is employed. The Fast Sockets library
140
receives the call made from the application
100
through the sockets API
90
to execute the emulation processing
141
to communicate using high-speed communication. For this reason, it is possible to take advantage of the high performance of high-speed communication while keeping application compatibility. As examples of Fast Sockets, there is the method disclosed in JP-A-11-328134, the method by Berkely University (refer to the paper by S. H. Rodrigues, T. E. Anderson, D. E. Culler, “High-Performance Local Area Communication With Fast Sockets”, Proceedings of the USENIX'97, 1997, pp. 257 to 274), the method by Shah et al. (refer to the paper by H. V. Shah, C. Pu, R. S. Madukkarumukumana, “High Performance Sockets and RPC over Virtual Interface (VI) Architecture”, Proceedings of CANPC'9, 1991), Winsock Direct made by Microsoft Corp. (refer to the article “Winsock Direct Specifications, on the Microsoft Windows Driver Development Kit (DDK)”.
When data
102
of the application
100
is registered (
800
) to perform communication, a processing overhead (
132
,
123
) of the buffer registration
800
occurs. When the data length is long, this overhead (
132
,
123
) is shorter than the communication time, so high communication performance is obtained. On the other hand, when the data length is short, this overhead is longer than the communication time, so the communication performance is reduced. In order to solve this problem, the Fast Sockets library
140
on its initialization allocates a pre-allocated buffer
142
and registers (
801
). When communicating short data
102
, this data is not registered, but is copied to the pre-allocated buffer
142
to perform the communication. In this case, despite the overhead of the copy, since the data length is short, and this overhead is small when compared to the registration processing, high performance can be obtained. While the pre-allocated buffer
142
is usually separated into buffers for sending and buffers for receiving data, these buffers are collectively shown in the form of one buffer
142
in FIG.
3
and the following figures of the software structure.
Above, the TCP/IP communication and the Fast Sockets have been described. While applications generally use TCP/IP communication (and as a result, the Sockets API), scientific computing applications use APIs such as MPI (Message Passing Interface Forum: refer to the standard “MPI: A Message-Passing Interface Standard”, 1995). Since MPI is independent of the computer architecture, when implementing MPI over high-speed communication, the calls made to the MPI API are mapped onto the calls of the high-speed communication API
91
. As an example of a product implementing this mapping, there is MPI-Pro made by MPI Software Technology Inc. (refer to the paper by R. Dimitrov and A. Skjellum., “Efficient MPI for Virtual Interface (VI) Arch
A. Marquez, Esq. Juan Carlos
Fisher Esq. Stanley P.
Hitachi , Ltd.
Lane Jack
Reed Smith LLP
LandOfFree
Method of transferring data between memories of computers does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method of transferring data between memories of computers, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of transferring data between memories of computers will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3363201