Multiple port shared memory interface and associated method

WO 95/05635

PCT/US94/09364

MULTIPLE-PORT SHARED MEMORY INTERFACE

AND ASSOCIATED METHOD

BACKGROUND OF THE INVENTION

1· Field of Invention

The invention relates generally to memory interface systems, and more specifically to a method and apparatus for providing and arbitrating access to a shared memory by multiple devices, for applications such as multiprocessor systems and communications switching.

2. Description of the Related Art

Shared memories are used to facilitate data passing 10 between multiple processes. A typical shared memory implementation involves the use of multiple ports. Each port may provide shared memory access to a different external device. The different devices, in turn, may be involved with the control or execution of different processes that pass the data.

In order to provide each port with access to the shared memory which is not significantly impacted by shared memory accesses by other ports, the path to the shared memory typically is designed with a bandwidth close to the sum of the bandwidths of the individual ports. This ensures that the data carrying capacity of the path is sufficiently large so that no port suffers a significant delay in accessing the shared memory even though multiple ports may seek access to the shared memory. Generally, this is achieved in one of two ways, or some combination thereof. First, the shared memory access time may be designed to be much faster than the data transfer times for the individual ports. Second, the path width to the shared memory may be designed to be

WO 95/05635

PCT/US94/09364

<img file="CA2168666C_D0001.tif" />

—2 — much greater than the path widths of the individual ports.

The first approach is to assign a time slot to each port during which data can be read from or written to the shared memory. For each port, the assigned time slot is shorter than the actual time required to transfer data through the port. Thus, the data is buffered temporarily during its transfer between a port and the shared memory. Ordinarily, the length of the port time slot is inversely proportional to the number of devices sharing access to the shared memory. A device using a particular port can gain access to the memory only during the assigned time slot for that port. Data is buffered between time slots. There have been shortcomings with this approach. For example, it often requires the memory access time to be significantly less than the port data transfer time. However, it is often impractical to have such, a mismatch between data transfer rates for the ports and the shared memory.

The second approach also involves assigning a time slot to each port. For example, writes to the shared memory involve temporarily buffering multiple data words received at a respective port, and then providing them to the memory on a wide memory path all in one memory access cycle during the time slot designated for the port. Conversely, reads from the shared memory involve providing multiple words on the wide path all in one memory access cycle during a time slot designated for a respective port; temporarily buffering the words read from memory; and then transferring the words through the port. This second approach is particularly well suited to burst-mode systems in which data words are communicated in multi-word bursts through respective ports. A complete burst may be temporarily stored in a memory access buffer, and then may be written or read

WO 95/05635

216366Ô

PCT/US94/09364 —3— during a single memory access cycle through such a wide bandwidth path to the shared memory. By providing a path with an appropriately large bandwidth, each port may be made to appear to have exclusive access to the shared memory, unimpeded by data transfers through other ports.

The illustrative block diagram of Figure 1 depicts an earlier implementation of a multi-port memory system in which k ports, each having word width m, equally share a common memory. Each burst includes k words. K memory access buffers each can store k m-bit words. Each buffer is connected to the shared memory by a k x m line wide path. The shared memory is k x m bits wide.

The illustrative drawings of Figure 2 show a data format used in a typical multiport memory system such as that in Figure 1. During a transfer of data into the shared memory, a k word burst of m-bit words passes through a port. The entire burst is briefly stored in a single memory access buffer. Then, during a prescribed time slot, all k-words of the burst are simultaneously transferred from the buffer and written to the shared memory on the k x m path. During a transfer of data out of the shared memory, k words are read from the shared memory during another prescribed time slot and are transferred to a single memory access buffer. Then the buffered data is transferred through the port associated with that buffer.

The port which originally inputted the burst may be different from the port that outputs the burst. The shared memory temporarily stores the burst so that it can be routed from the input port to the output port. Thus, the system of Figure 1 can be used to pass data between ports.

WO 95/05635

PCT/US94/09364

216866g

-4More specifically, for example, in a memory write operation, k m-bit words received through a respective port are buffered by a memory access buffer assigned to that port· Subsequently, during a time slot reserved for that memory access buffer, all of the k m-bit words stored in the assigned buffer are simultaneously written to the shared memory on the shared k x m-bit wide path.

In a like manner, each of the other buffers can store m-bit words on behalf of their own associated ports. The entire contents (all k words) of each individual buffer can be written to the shared memory during the individual time slot reserved for that buffer. A memory read operation is analogous, with the steps of the write operation reversed.

A disadvantage of this prior implementation is the large number of interconnection pins between a buffer and the bus in a large multi-port system. Figure 3 is a block diagram depicting another conventional multi-port shared memory system. The data format used involves 16 word bursts of 72 bits per word (64-bit data plus 8-bit parity) . The shared memory bus has a width of 1152 lines (16 words x 72 bits/word). The bus is connected to each of 16 memory access buffers. Each buffer would need more than 1224 data pins, 1152 to connect to the bus and 72 to connect to the port. Unfortunately, the 1152 data pins connected to the bus would each require high drive capability to operate on a bus connected to all 16 buffers and to the shared memory. Figure 3 shows an illustrative bus capacitance that must be overcome by the buffer pins.

Thus, there has been a need for an improved architecture for providing multiple port access to a shared memory. The architecture should require fewer pins for memory access buffers and should not require

WO 95/05635

2.168666

PCT/US94/09364 —5— high drive capability for buffer pins. The present invention meets those needs.

SUMMARY OF THE INVENTION

A novel multi-port memory system is provided. The memory system includes a random access memory. Multiple buffers are provided for temporarily storing data during transfer of the data between respective ports and the random access memory. A data path interconnects the random access memory and the multiple buffers. An interconnect circuit conducts different prescribed subsets of data between the ports and different prescribed buffers.

In another aspect, a method is provided for the transfer of port bursts between a random access memory and multiple data ports. A port burst is a multi-word burst of data transferred through a single port. Data words in respective port bursts are partitioned into multiple subsets. Each data word subset is associated with a temporary storage buffer. Data word subsets are conducted between respective ports and their respective associated buffers where they are temporarily stored. An entire port burst can be transferred between the multiple buffers and the random access memory in a single memory access cycle.

The present invention reduces the number of pin connections required to transfer a multi-word burst of data between a port and a shared memory. Further, a present embodiment of the invention reduces the output drive requirements of memory access buffers.

These and other purposes and advantages of the present invention will become more apparent to those skilled in the art from the following detailed

WO 95/05635

PCT/US94/09364

-6description in conjunction with the appended drawings in which:

BRIEF DESCRIPTION OF THE DRAWINGS 5 Figure 1 is a block diagram of an earlier multi-port shared memory system;

Figure 2 is a data format which can be used in the earlier system of Figure 1;

Figure 3 is a block diagram of another earlier multi-port shared memory system;

Figure 4 is a generalized block diagram of a first multi-port shared memory system in accordance with the present invention;

Figure 5 shows a data format and data flow used in the embodiment of Figure 4;

Figure 6 is a more detailed block diagram of memory access buffers and control logic of the embodiment of Figure 4;

Figure 7 is a block diagram of a representative dual register pair of the memory access buffers of Figure 6;

Figure 8 is a timing diagram which explains the operation of the control logic and memory access buffers of Figure 6;

Figure 9 is a timing diagram which explains the operation of the dual register pairs of Figure 7; and Figure 10 is a block diagram of an Asynchronous

Transfer Mode switch in accordance with the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The present invention comprises a novel method and apparatus for implementing a multi-port shared memory system. The following description is presented to enable any person skilled in the art to make and use the invention. Descriptions of specific applications are provided only as examples. Various modifications to the preferred embodiment will be readily apparent to those

WO 95/05635 i

-7PCT/ÜS94/09364 skilled in the art, and the generic principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the invention. Thus, the present invention is not intended to be limited to the embodiment shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

Referring now to Figure 4, there is shown a block 10 diagram of a first multi-port shared memory system 18 in accordance with the present invention. The first system 18 includes a set of k ports 20 each inputting/outputting m-bit data words in n-word bursts, an interconnection matrix circuit 22, a set of m memory access buffers 24, and a shared memory 26. A port serves as a two-way digital path on which digital information can be transferred to or from external circuitry such as a data bus. Port structures are well known to those skilled in the art and need not be described herein. Specific regions within the shared memory 26, indicated by dashed lines, are reserved for subsets of the data temporarily stored in the buffers 24.

During memory write operations, the ports 20 transfer binary data from individual external devices (not shown) connected thereto. These ports 20 provide the received binary data to the interconnection matrix 22, which distributes the data transferred through the ports among the memory access buffers 24. In the present embodiment, the distribution is accomplished such that each memory access buffer 24 receives a subset of the data transferred by each of the ports. In the present embodiment, each buffer 24 can transfer in parallel to memory 26 all of the data received by the buffer from an individual port. Moreover, all of the parallel transfers of data received from such an individual port can occur during a single memory access cycle.

WO 95/05635

<img file="CA2168666C_D0002.tif" />

PCT/US94/09364 φ

-8The memory read operation is analogous with the steps of the write operation, but reversed. That is, the bits for n m-bit words are transferred from the shared memory and are distributed among the memory access buffers 24. The interconnection matrix 22 then provides the distributed bits to a single port 20 through which a burst comprising n m-bit words is transferred to an external device (not shown).

Referring now to Figure 5, a diagram is shown which illustrates details of the interconnection matrix circuit 22, the data format and data flow presently used in the first system 18. Each memory access buffer 24 is dedicated to storing a specific bit position for every data word transferred through any of the ports. Data words may contain data and parity information. In a presently preferred embodiment, the total number of memory access buffers is equal to the total number of bits per word (m), thus allowing each buffer to be responsible for a single bit position in every word. For example, memory access buffer 24-1 stores bit 1 (BI) of each word transferred through any of the ports 20-1 to 20-k; memory access buffer 24-2 stores bit 2 (B2) of each word transferred through any of ports 20-1 to 20-k; etc.

Referring again to Figure 4, each memory access buffer 24-1 through 24-k is connected to the shared memory 26 by a set of n data lines 28-1 through 28-k. Thus, the m-subsets of different order bits can be simultaneously transferred between the memory 26 and the buffers 24. In the present embodiment, there are n-bits for each of the m-subsets; that is, one bit for each of the n words in an n-word burst. Specifically, if there are n-words per burst, and there are m-bits per word then there are n-bits per burst of order BI, n-bits per burst of order B2, ..., and n-bits per burst of order Bm. All n-bits of a prescribed order can be simultaneously

WO 95/05635

PCT/US94/09364

-9provided to the memory 26 on the n lines connected to a buffer assigned to store the bits of that order. For example, all n of the BI bits are provided on the n lines connected to buffer 24-1. This means that the buffers can transfer n words at a time to or from an addressed location of memory. As explained below, all of these transferred words would have originated at or been targeted to the same external device.

Each port 20 transfers data in a prescribed format between the interconnect matrix 22 and external circuitry (not shown) . In this document, a port burst shall mean a data burst transferred through a single port in either direction to or from the external circuitry. In a present embodiment as shown in Figure 5, m-bit data words may be presented by a port to the interconnect matrix 22 for transfer to the buffers 24. Conversely, m-bit data words may be presented to such a port by the interconnect matrix 22 for transfer to the external circuitry. The data word format involves data words with bits BI through Bm presented in a prescribed order. The interconnect matrix distributes the bits as described above.

The interconnect matrix circuit 22, for example, can be implemented using a printed circuit board (PCB), wire wrap or soldered wires. In the currently preferred embodiment, the interconnect matrix provides the connections between ports and buffers. As explained below, for example, each BI in each word transferred through port 20-1 is connected by the matrix 22 to a storage element of buffer 24-1 associated with port 20-1; each BI in each word transferred through port 20-2 is connected by the matrix 22 to a storage element of buffer 24-1 associated with port 20-2; ... each BI in each word transferred through port 20-k is connected by the matrix 22 to a storage element of buffer 24-1 associated with port 20-k.

WO 95/05635

PCT/US94/09364

-10The memory access buffers can be implemented using either standard or custom logic. The drawings of Figures 6-9 illustrate and explain the operation of the memory access buffers 24 and a representative dual register pair 34 of one of the memory access buffers 24. The illustrative block diagram of Figure 6 shows the memory access buffers 24, control logic 29, serial input/output lines 30 connecting the buffers 24 to an interconnect matrix (not shown) and a n-bit parallel input/output bus

32. The control logic 29 controls serial I/O on lines 30 and parallel I/O on bus 32. The operation of the control logic will be explained with reference to the timing diagram of Figure 8.

Referring to Figure 7, there is shown a representative dual n-bit register pair 34. In a present embodiment of the invention, each memory access buffer 24 includes k dual register pairs like the one shown in Figure 7 for a total of 2k registers per buffer 24. Each register pair in each buffer is associated with one of the ports. Each register pair is dedicated to storing all bits in a prescribed location (order) in every data word transferred to or from the port associated with that register pair. That is, each respective dual register pair of a respective buffer stores and transfers all of the same order bits from each data word transferred through a respective port.

For example, referring to Figure 5, in buffer 24-1, dual register pair R^ receives and stores the first (in order) bit BI of each data word transferred through port Pi (not shown). In buffer 24-2, dual register pair R^ receives and stores the second (in order) bit B2 of each word transferred through port Pj. In buffer 24-m, dual register pair R^ receives and stores the 1¾ (in order) bit B_m of each word transferred through port P_{. Thus, for example, all Bis transferred through port Pi represent

WO 95/05635

PCT/ÜS94/09364

21686GÛ —11— a subset of the data transferred through that port. Similarly, all B2s represent another subset as do all B_ms, for example.

The dual register pair 34 includes a shift-in register 36 and a shift-out register 38. The shift-in register 36 first stores and then asserts in parallel n bits of data onto the memory bus 32 to be written into the shared memory 26. The shift-out register reads in parallel and stores n bits of data from the memory bus 32 which have been read out of the shared memory 26.

The data stored in the shift-in register 36 is serially shifted from the interconnect matrix 22 into the shift-in register 36; and from there, it is asserted in parallel onto the bus 32 to be written into the memory 26 as described above. Conversely, the data stored in the shift-out register 38 is read in parallel from the bus 32 into the shift-out register after it has been read from the memory 26; and from there, it is serially shifted from the shift-out register to the interconnect matrix

22. Output enable buffer 40 controls the serial transfer of data from the shift-out register 38 to the interconnect matrix 22.

Referring to Figure 8, there is shown a timing diagram for the memory access buffers of Figures 4-6.

The timing diagram of Figure 8 depicts the timing for a thirty-two port system (k=32), with each port having a

0 burst size thirty-two of words (n=32) . Thus, n=k. In a present embodiment, memory access alternates between 32 memory read cycles followed by 32 memory write cycles.

Focusing first on the timing of memory writes, during the Memory Data Write period (clock ticks 33-64), a different thirty-two word burst of data can be written to the common memory 2 6 for each port 20. Each port has

WO 95/05635

PCT/US94/09364 —12 — a clock cycle during which data transferred through that port can be written to memory 26. Consecutive ports' data are written on consecutive clock cycles, which are the respective write time slots for the respective ports. In order to prepare for the write time slot allocated to it, each port can begin transferring data to the memory access buffers thirty-two clock cycles before its time slot, so that there is sufficient time to transfer all thirty-two words of a burst received by a port to the buffer assigned to the port before the arrival of the port's time slot. Thus, during the thirty-two clock cycles preceding the wl.C on the Memory Data timing line, there are thirty-two transfers from port 20-1, via the interconnect matrix 22, to the memory access buffer 24-1, depicted as wl.l, wl.2 ... wl.32 on the port 20-1 Data line. Since each port has a different time slot, the transfers from the ports are appropriately staggered so that each port's burst is ready to transfer to the shared memory just prior to the port's time slot.

More specifically, for example, during each Port 20-1 Data Write Cycle shown in Figure 8 all of the bits in a first word transferred through Port 20-1 are distributed by the interconnect matrix 22 among the assigned, dual port register pairs of the memory access buffers 24. For example, assuming that there are thirtytwo words per burst and m bits per word, then during the first Port 20-1 Data Write Cycle, (wl.l) bit 1 (BI) of word 1 (Wl) transferred through Port 20-1 is serially written to a prescribed register pair in buffer 24-1; B2 of Wl is serially written to a prescribed register pair in buffer 24-2; B3 of Wl is serially written to a register pair in buffer 24-3; etc. Bm of Wl is serially written to a register pair in buffer 24-m.

WO 95/05635

21G8666

PCT/US94/09364 —13 —

Similarly, for example, during the 32nd Port 20-1 Data Write Cycle (wl.32), BI of W32 transferred through port 20-1 is serially written to the same register pair of port 24-1 as was BI of W1 transferred through port 20—1. Likewise, for example, during the third Port 20-1 Data Write Cycle (wl.3), B6 of W3 transferred through port 20-1 is serially written to the same register pair in memory access buffer 24-6 (not shown) as was B6 of W1 transferred through port 20-1.

Thus, during each of the thirty-two Data Write Cycles for port 20-1 (wl.l through wl.32), data is serially shifted into designated shift-in registers in the respective memory access buffers 24-1 through 24-m. During Memory Data Cycle wl.C, all of the data that has been serially shifted into the shift-in registers of port 20-1 during the thirty-two prior serial shift—in cycles (wl.l through wl.32) are asserted in parallel onto the bus 32 so they can be written into the shared memory 26.

Similarly, during the Memory Read period (clock ticks 1-32), thirty-two words of data can be read from the shared memory for each port. For example, during a single one clock memory access cycle at rl.C, thirty-two words can be read from memory 26; these words can be transferred to an external device through port 20-1.

Thus, rl.C is the read time slot for port 20-1. Likewise, for example, r31.C represents the read time slot for port 20-31.

Specifically, for example, at time slot rl.C a shift-out register in buffer 24-1 receives in parallel all Bis of the thirty-two words read in parallel from the memory 26. Similarly, for example, at time slot rl.C a shift-out register in buffer 24-18 (not shown) receives in parallel all B18s of the thirty-two words read in parallel from the memory 26.

WO 95/05635

PCT/US94/09364

-14During the thirty-two port 20-1 Data Read Cycles (rl.l through rl.32) the words read from the memory 26 during Memory Data Read Cycle (rl.C) are serially transferred, via the interconnect matrix 22, to port 20-1. For example, during Data Read Cycle (rl.l), all of the bits for the first word (Wl) to be transferred out through port 20-1 are simultaneously serially shifted from m different shift-out registers in m different buffers 24, via the matrix 22, to provide Wl to port 20-1. Likewise, for example, during Data Read Cycle (rl.31), all of the bits for the thirty-first word (W31) to be transferred out through port 20-1 are simultaneously serially shifted from m different shift-out registers in m different buffers 24, via the matrix 22, to provide W31 to port 20-1.

Thus, when data received at a prescribed port is to be written into memory 26, the interconnect matrix distributes the data word bits among multiple buffers so that each buffer only stores a subset of each word to be stored. In the preferred embodiment each buffer stores only bits of a prescribed order. Thus, each buffer need only provide to the bus 32 a subset of all of the data to be simultaneously transferred to memory.

Conversely, when data read from memory is to be output to a prescribed port, the interconnect matrix recombines the bits distributed among multiple buffers so that they are outputted as a series (or burst) of complete words. The entire data to be transferred out through a port is outputted in parallel from the memory and is inputted to the buffers. However, each buffer only receives a subset of that entire data. The interconnect matrix 22 recombines the subsets into a burst of data words.

WO 95/05635 —15—

PCT/US94/09364

The illustrative timing diagram of Figure 9, shows the operation of the dual register pair of Figure 7. During the p.WRITE pulse, data that has been read from the shared memory 26 during an (rX.c) Memory Data Cycle is written in parallel into the shift-out register 38. During each SHIFT-OUT pulse, a single bit is shifted-out from the shift-out register to the interconnect matrix 22 for transfer to a respective port. For example, for buffer 24-1, each SHIFT-OUT pulse corresponds to an rl.X pulse in Figure 8. During each SHIFT-IN pulse, a single bit of data is shifted in to the shift-in register. For example, for buffer 24-1, each SHIFT-IN pulse corresponds to an wl.X pulse in Figure 8. During the P.READ pulse, all of the data that has been shifted in to the shift-in register can be read in parallel to be written to the shared memory 26 during a wl.C cycle.

In the present embodiment, the shared memory 26 can be a standard random access memory (RAM) configuration with a memory width of (n x m) . Thus, the memory is able to transfer (n x m) bits of data to or from the memory access buffers 24 simultaneously, and is able to store all (n x m) bits in a single addressed location of memory, as depicted in Figure 5. In the present embodiment, rather than store each word from a port as contiguous bits, the words are interleaved,¹¹ so that all the first order = order bits (Bis) are stored contiguously, followed by all of the second order bits (B2s), and so forth. In this configuration, all of the bits representing the n words from or to a specific port 20 can be readily stored as a line of the shared memory 26.

In the present embodiment, time slots are preassigned. However, time slots could be arbitrated based on priorities. Moreover, since in the current embodiment each of the k ports 20 requires a single clock cycle time slot to communicate with the shared memory 26,

WO 95/05635

PCT/US94/09364 ”16 and each port requires n clock cycles prior to that time slot to transfer an entire n-word data burst to or from the memory access buffers 24, it is efficient to set the number of words per burst (n) equal to the number of ports (k), thereby providing a smooth, cyclical process for transferring data without bottlenecks or idle times. The relationship between the number of ports and words per burst can change, however, without departing from the invention.

Figures 10 is a block diagram of 32-port Asynchronous Transfer Mode (ATM) switch in accordance with the present invention. ATM is a payload multiplexing technique for information transfer using fixed-size packets, called cells. In a current implementation, an ATM cell is 53-bytes long and consists of a 5-byte header which carries the routing information, followed by a 48-byte information field (payload). The payload of each ATM cell entering the switch is placed in a specific location in the shared memory. As the ATM cell payload is 48-bytes wide (8 bits per byte), shared memory is made 384-bits wide (48 x 8 = 384). This implementation was selected to have a 3 2-bit bus from the memory access buffers to the memory with each port word 12-bits wide, and with port bursts of thirty-two words. Note that 384 is also equal to (12 x 32); that is, a cell is transferred in a 32-word burst of 12-bit words.

The ATM switch routes each received cell on a port to a destination port according to routing information that is included in each such cell. More particularly, a cell is transferred from a port and stored in shared memory. Then it is retrieved from the shared memory and is transferred to a destination port indicated in the cell. In this manner, data can be switched between ports. The operation of the control memory and the switch controller will be understood by those skilled in

WO 95/05635

PCT/US94/09364 the art, form no part of the invention and need not be described herein.

While particular embodiments of the invention have 5 been described in detail herein, it will be appreciated that various modifications can be made to the preferred embodiment without departing from the scope of the invention. Thus, the foregoing description is not intended to limit the invention which is defined in the appended claims.