Determining a message residue

A variety of computer applications operate on messages to create a message residue. The residue can represent message contents much more compactly. Among other uses, message residues are frequently used to determine whether data transmitted over network connections or retrieved from a storage device may have been corrupted. For instance, a noisy transmission line may change a “1” signal to a “0”, or vice versa. To detect corruption, a message is often accompanied by its message residue. A receiver of the data can then independently determine a residue for the message and compare the determined residue with the received residue.

A common message residue is known as a Cyclic Redundancy Check (CRC). A CRC computation is based on interpreting a stream of message bits as coefficients of a polynomial. For example, a message of “1010” corresponds to a polynomial of (1 x 3 )+(0 x 2 )+(1 x 1 )+(0 x 0 ) or, more simply, x 3 +x 1 . The message polynomial is divided by another polynomial known as the modulus. For example, the other polynomial may be “11” or x+1. A CRC is the remainder of a division of the message by the polynomial. CRC polynomial division, however, is somewhat different than ordinary division in that it is computed over the finite field GF(2) (i.e., the set of integers modulo 2). More simply put: even number coefficients become zeroes and odd number coefficients become ones.

FIG. 1 is a diagram illustrating determination of a message residue.

FIG. 2 is a diagram illustrating operations to determine a message residue.

FIG. 3 is a diagram illustrating table lookups that can be used in determining a message residue.

Determining residue of a message occurs frequently in a variety of applications. This computation often represents significant overhead in processing a message, for example, for transmission or after receipt. The following describes techniques that can speed this computation by permitting simultaneous processing of different portions of a message.

To illustrate, FIG. 1 depicts CRC computation of a message M 100 of length 3 N-bits. The bits of the message M represent coefficients of a polynomial M(x). As described above, a CRC is the modular remainder of the message, M(x), with respect to a polynomial P(x) modulus (“poly”):

CRC M ( x )= M ( x )mod P ( x )  [1]

where P(x) is another polynomial represented by a bit-string. In practice, P(x) is defined for different applications. For example, iSCSI (Internet Small Computer System Interface), which uses CRC for end-to-end data protection, uses the value of 11EDC6F41H 16 as the polynomial P(x). Other applications have selected different polynomial sizes and values. For a typical k-bit CRC residue, P(x) will be a k-bit polynomial (e.g., 32-bits) and M will be a message m(x) shifted k-bits. The k-bit CRC value is typically stored in the vacant, least significant k-bits of M(x).

A CRC value reflects the value and position of each bit in a message M. However, as shown in FIG. 1 , the CRC value may be computed in parallel by independently processing segments of the message and later combining the results in a way to preserve proper CRC computation. More precisely, the CRC of M(x) can be broken down into the computation of the CRC for different segments (A, B, C) of the message. For example:

CRC M ( x )=(CRC( C )· K 2)+(CRC( B )· K 1)+(CRC( A ))  [2]

where “+” represents polynomial addition (e.g., bitwise XOR) and “·” represents polynomial multiplication (e.g., carry-less multiplication of bit strings mod P(x)). The values of CRC (C), CRC (B), and CRC (A) can be independently and simultaneously computed to speed computation of the overall CRC value. The CRC values for each segment can then be combined into the overall CRC value of message M by polynomial multiplication, e.g., by constants K. For example, in the example shown in FIG. 1 which features uniform segments of size N-bits, K 2 =x 2N mod poly and K 1 =x N mod poly where x 2N and x N correspond to the significance of the segments within the message. Since x 2N and x N are constants, and the modulus polynomial poly is constant, the values of K 2 and K 1 are, therefore, also constants that can be pre-computed before the value of message M is even accessed. That is, the values of constants, K, depends on the segmentation points and known value of poly, rather than the bit values of message M. In other words, a given set of constants determined for a given polynomial can be used for different messages. Sets of constants may be pre-computed and provided for the different polynomials used in practice. The values of the constants can be determined using a CRC operation. For example, in the above example, K 2 can be computed as CRC (x 2N ) for a specified polynomial.

While FIG. 1 depicted three segments, other implementations may use different segmentation schemes using a different number of segments, segments of non-uniform size, and so forth. Additionally, while the values of these constants may be pre-computed, they may also be computed concurrently with the segment CRC values (e.g., during initial processing of a first message or, less efficiently, for each message processed).

The potential advantages of such a scheme can be quantified in different environments and implementations. For example, a processor may feature a dedicated CRC instruction (e.g., a CRC or other message residue processor macro-op). Such an instruction may have a syntax of:

CRC (input-bit-string, message-residue)

where the input-bit-string (e.g., 64, 32, 16, 8, or other specified or hard-wired string length) can represent a portion of a message. Optionally, the instruction or a companion instruction may specify the polynomial or bit-string-width. The instruction can update the CRC message-residue value to reflect the incremental contribution of the new input-bit-string. Using such an instruction, computing the CRC of a message M can be done cumulatively, using a sequence of calls to the CRC instruction that consumes the message in input-bit-string sized chunks with each invocation of the CRC instruction incrementally updating the CRC value. After all data chunks are consumed, the remaining message residue reflects the CRC value of the message as a whole.

Such an instruction could be implemented by the following processor micro-ops which implement a 32-bit CRC operation:

TEMP1[63-0]←BIT_REFLECT64 (SRC[63-0]) TEMP2[31-0]←BIT_REFLECT32 (DEST[31-0]) TEMP3[95-0]←TEMP1[63-0]<<32 TEMP4[95-0]←TEMP2[31-0]<<64 TEMP5[95-0]←TEMP3[95-0]XOR TEMP4[95-0] TEMP6[31-0]←TEMP5[95-0]MODULO POLY DEST[31-0]←BIT_REFLECT (TEMP6[31-0]) DEST[63-32]←00000000H

where BIT_REFLECT changes the endian representation of a string if necessary.

In the sample implementation to determine a CRC value described above, each CRC instruction required the output of a previous CRC instruction as input. This serial approach, however, consumes a number of processor cycles that is at least T*L where T is the CRC instruction cycle latency and L is the number of chunks. By contrast, computing the segments in parallel results in a latency of approximately (T*L/NumSegments) plus the latency of the recombination operations which is, by comparison, relatively insignificant. Generally, the later approach will represent a significant decrease in the time used to determine a CRC.

FIG. 2 illustrates a sample implementation that uses a processor provided CRC instruction. In the example depicted, the particular CRC instruction takes three cycles to complete. The timing diagram illustrates that the processor implements a multiple-instruction execution pipeline.

As shown, the CRC instruction operates on a sub-segments (e.g., A[i], B[i], and C[i] where i=1 to 3) of each segment in succession. The residue for each segment is stored in an incrementally updated residue variable. For example, the residue value, X, is updated for each sub-segment chunk (e.g., C[ 1 ], C[ 2 ], C[ 3 ]) of segment C. The residue values may be initialized to values appropriate to a given application (e.g., all 1-s for iSCSI).

Due to the pipeline architecture and latency of the CRC instruction, operations on chunks of a given segment are interleaved with operations on chunks of other segments. For example, as illustrated, the like-positioned chunks of segments A, B, and C (e.g., [A[ 1 ], B[ 1 ], and C[ 1 ]) are sequentially inserted 104 a - 104 c into the pipeline. Once the segment residue for a given chunk has been updated, the next chunk for that segment can be processed. For example, after completion of CRC instruction 104 a for chunk C[ 1 ], CRC instructions 104 d can process chunk C[ 2 ]. The illustrated overlap between CRC instructions represents an overall latency reduction attributable to techniques described above.

As shown, after determining X, Y, and Z, recombination 106 proceeds. Potentially, recombination 106 can overlap the CRC computations. For example, determining X·K 2 may be initiated as soon as the value of X is determined instead of waiting for the values of Y and Z. Potentially, the message residue values of X, Y, and Z may be stored in the message itself (e.g., the value of X may be stored in C[ 3 ], Y in B[ 3 ], and Z in A[ 3 ]) changing the message into sparse storage of the message residues.

While FIG. 2 illustrated a particular implementation for a given environment, many variations are possible. For example, computation of the segment CRC values may be determined by different processor elements (e.g., programmable processor cores of a processor integrated on a single die) in addition to or in lieu of pipeline processing. In addition, the number cycles of a CRC instruction may be larger or smaller than the 3-cycle implementation depicted. Potentially, the number of message segments may be based on (e.g., equal to) the number of cycles used by a CRC instruction. For example, a 4-cycle instruction may be paired with an implementation that divides messages into 4-segments and process the 4-segments simultaneously. Further, while a dedicated CRC instruction was depicted, the techniques may also be applied where no explicit dedicated CRC instruction is provided by a processor. Additionally, again, while a CRC computation was illustrated in FIG. 2 , techniques described above may be used in other message residue determinations.

Additional variations can further or independently speed computation of a message residue. For example, the approach illustrated above computed (CRC(segment)·K). A straight-forward implementation can simply perform this polynomial multiplication and modular reduction after determining a segment's CRC value. However, since K and poly are constants, this operation can instead be implemented using a set of lookup tables. For example, as shown in FIG. 3 , modular polynomial multiplication of a segment's CRC value (denoted Y) 108 by a constant K can be expressed as:

Y·K=Y 3 ( D 3 ·K )mod poly+ Y 2 ( D 2 ·K )mod poly+ Y 1 ( D 1 ·K )mod poly+ Y 0 mod poly  [4]

where Y i represent n-bit segments of Y, and D i represent constants of x Di . The values of (D i ·K) represent the modular polynomial multiplication of two constants and, therefore, are also constant. Thus, instead of actually computing Y i (D i ·K) mod poly, each value of Y i can be used as an n-bit lookup key into a table 110 a , 110 b , 110 c storing pre-computed possible values of Y i (D i ·K) mod poly. For example, where Y is a 32-bit number, table 110 a may store pre-computed values of Y 3 (D 3 ·K) mod poly for lookup by the 8-bit value of Y 3 . Such tables can speed these computations considerably without consuming excessive amounts of data storage. Additionally, such lookups may be performed in parallel. The resulting lookup values can be quickly summed using polynomial addition.

The above described CRC message residues. Such residues can be stored Random Access Memory, for example, in a packet stored in memory or for comparison with residues stored in received packets, for example, to determine whether data corruption has likely occurred. The techniques described above are not limited to CRCs, however, but may be used in other schemes, such as other coding/error correction schemes that operate on modular residue, for example, in a finite field of polynomials over GF( 2 ). Additionally, while the above techniques repeatedly performed modular reduction on intermediate values, such modular reduction may be deferred or performed at other junctures.

The techniques described above may be implemented in a variety of logic that includes hardware, software, and/or firmware or a combination thereof. For example, techniques described above may be implemented as instructions of a computer program disposed on a computer readable medium that causes a processor to perform the message residue determination described above.

Other embodiments are within the scope of the following claims.