Method and Apparatus for DRAM Memory Performance Enhancement

The present invention relates to computer memory systems, and in particular to cache memory systems.

High speed computer memories, in particular cache memory performance is strongly influenced by the amount of time necessary to fill the cache memory with new data required by the processor. The sooner such a read operation can be started in main memory, the less time it will take to fill the cache. The influence of read latency on performance is derived from the read latency times the cache miss rate. This is the amount of time by which an average processor cycle will be increased due to cache read latency.

The method and apparatus according to the present invention starts a read operation before a previously issued write operation, except that a read operation need only wait for write operations to complete if they have an impact on the data to be obtained by the read operation. Therefore a read may be promoted ahead of any write operation which does not conflict.

The opportunity for promoting reads over writes occurs anywhere writes are buffered. In these places the read operation address range is compared to older write operations stored in the buffer or queue. If the addresses do not overlap, then the write can have no effect on the result of the read and the read is promoted and issued first. If the addresses do overlap, then the read waits for the write to be issued. The address comparison may be done sequentially by stepping through the queued write commands, as in this implementation, or it may be done in parallel by replicating the address comparison logic.

In a read cycle, the cycles for a typical static column DRAM are; (1) Row Address applied, (2) Row Address Strobe asserted, (3) Column Address applied, (4) Column Strobe asserted, (5) Strobes deasserted for a precharge time. If the Row address is the same from one operation to the next, then phase 5 of the current cycle and phases 1 and 2 of the next cycle (and 4 for static column devices) may be omitted thereby decreasing the amount of time required for the next cycle. Identical savings may be made with write cycles. Address comparison logic is used to determine if the bits which select the set of DRAMs to be used and the Row Address Strobe address bits are the same. This ability to use a shorter cycle increases memory performance and reduces the amount of time it takes to clear writes from the command queues and also reduces the latency on read operations.

These and other features of the present invention are better understood by reading the following detailed description of the invention, taken together with the drawings, wherein:

FIG. 1 is a prior art DRAM control state diagrams;

FIG. 2 is a DRAM control state diagram according to one embodiment of the present invention;

FIG. 3 is a block diagram of one embodiment of a DRAM address and control unit according to the present invention; and

FIG. 4 is a block diagram providing further detail of the command and address range elements showing sequential address comparisons.

Further details of one embodiment of a memory system is shown in copending application U.S. patent application Ser. No. 07/263,112 A FLOATING POINT PIPELINE PROCESSOR, filed Oct. 26, 1988, incorporated by reference, and U.S. patent application Ser. No. 07/264,058, entitled AN EFFICIENT CACHE WRITE TECHNIQUE THROUGH DEFERRED TAG MODIFICATION, filed concurrently herewith and incorporated by reference.

The following memory device read or write steps, cycles, states, signals and associated hardware, individually, and in the prior art combination, are known to those of ordinary skill in the art and further defined in product specifications and application notes provided for exemplary DRAMS 256kx4 100 ns static columns DRAM manufactured by Toshiba TC514258J-10, Mitsubishi M5M44C258J-10, Hitachi HM514258JP-10 and FUJITSU MB81C4258-10PJ.

In the prior DRAM cycle illustrated in the state diagram 50 of FIG. 1. When a standard cycle starts (from Idle, 52), the DRAM row address portion of the command is driven on the DRAM address lines and the DRAM control lines, Row Address Select (RAS) Chip Select (CS) and write W are deasserted. The RAS control line is asserted in state RAS 54, then, in state CS/W the columns address is driven on the DRAM address lines and either CS (for a read cycle) or CS and W (for a write cycle) is asserted. The normal progress is then to state data available 58 (DAV) to latch in the data on a read cycle, or to assert CS to perform a write cycle. The cycle then completes by proceeding to state Pchg, 60 which deasserts all control lines and then to Idle 52.

By contrast, the operation of the DRAM according to the present invention is illustrated in the state drawing 70, FIG. 2 and associate control 100 of FIG. 3. The operation of states 72, 74 and 76 correspond to states 52, 54 and 56 of FIG. 1. If, in DAV 78, we recognize the RAS address for the current cycle is the same as the RAS address for the next command to be executed, then we can branch back to CS/W 76 either directly or through state Ovlp 77. In the state Ovlp 77, both W, 123 and CS, 122 control signals are deasserted to avoid unintended write operations to the DRAM(s) 130. Ovlp 77 may be used all the time if the timing requirements of the DRAM cannot be met otherwise, but is always required when the current cycle is a read and the next cycle is a write so that there is no possible overlap between the assertion of CS for the previous cycle and W for the next cycle. The transition from state DAV 78 to CS/W 76 eliminates either two or three (depending on whether Ovlp is used) of the five states required for the standard cycle.

Similar savings accrue for a read-modify-write cycle. The standard (prior art) cycle performs a read cycle as described above passing through states Idle 52, RAS 54, CS/W 56 and RW1 62. Then the data read is modified and written to the same address in states RW2 64 and DAV 58. The cycle then completes by passing through Pcgh 60 and returning to Idle 52.

By contrast, the present invention provides that if, in state DAV 78, (having arrived through states 72, 74, 76, 82 and 84 analogized to previously described in reference to states 52, 54, 56, 62 and 64), the RAS address of the next command and the RAS address of the current cycle are equal, then the state machine passes to CS/W either through Ovlp 77 or directly 79. Therefore, the system and method of the present invention saves two or three (72, 74 and 80) of the seven cycles required by the standard cycle 50 of FIG. 1.

The memory structure 100 of FIG. 3 provided commands for DRAM operations (read, write, read-modify-write) which are entered into the command queue (110). At each cycle, the command progresses through the queue if the next adjacent entry is empty. At each level a comparison is done in parallel (102, 104, 106, 108) between the RAS address field of each entry in the queue and the RAS address of the current cycle being executed by the DRAM control (120). The comparator output associated with the next command to be executed by the DRAM control is selected (112) and presented to the DRAM control to affect the flow of its state machine. The DRAM control presents RAS, CS and W control signals 121, 122 and 123 and a multiplexed RAS/CS address on leads 114 to the DRAM array (130) to control its operation according to the state shown in diagram 70 of FIG. 2.

An embodiment of the control logic shown in FIG. 4 selects the next read command to be executed using multiplexer 210 and sequences through all older write commands in the queue using multiplexer 211 comparing the address ranges using comparitor 212. If any older write conflicts in address, then the control logic waits until that write command is being executed to promote the read to be the next operation to be executed. Otherwise the read command is promoted immediately and presented to the DRAM control logic via multiplexer 209 when the current command being executed (if any) is no longer needed by the DRAM control.

Moreover, the present invention permits the merging of the data from the read operation of the memory with data which is to be updated by the pending write commands in order to allow promotion of reads over write operations which do conflict in address.

Furthermore, the present invention when implemented can provide a sequence of commands to a processor permitting out-of-order execution of commands (including writes) in the main memory, optimized using these shorter cycles. For example, in the case of two operations which share a RAD address and are separated by other operation(s), the second operation could be executed out of order. ##SPC1##