Our aim in producing an ATM `internet', with ATM delivered into the end system, means that our consideration of adaptation is concerned primarily with efficient implementation in such end systems rather than as a means of interconnection of current packet or circuit switched traffic types. However, by providing end-to-end service for cell payloads, MSNL can support any adaptation layer, including those recommended by CCITT.
Within MSNA, and in keeping with the desire to eradicate unnecessary multiplexing, we have implemented only adaptation layers which do not perform cell level interleaving of different segmented blocks over a single virtual circuit. When cell level interleaving is required, distinct VCIs are used; that is, we use multiple MSNL streams and allow the VCI to perform its intended role of distinguishing them. In B-ISDN terms, this is equivalent to using a virtual path service where the end-to-end VCI bits distinguish cell interleaved blocks, rather than using a single VCI and sorting cells into blocks by yet another layer of multiplexing, for example by use of the MID in AAL-3/4.
In the MSNA architecture, the block level service is known as MSSAR
(multi-service segmentation and reassembly) and is built on top of
MSNL. We have been using a particular adaptation protocol for a number
of years [9]
, while investigating issues in
adaptation. We investigated the minimal information required to
implement sufficiently secure yet efficient and simple to implement
adaptation layers, and have now arrived, together with other
researchers, at a class of protocols which require that the ATM cell
header include a logical end-to-end user data bit. The bit can be used to
implement a range of different adaptation layers, but in this paper,
we concentrate on adaptation layer protocols for data services.
For data services, as a starting point we may take the definition:
The fundamental role of the data adaptation layer is, at the transmitter, to take a PDU and provide segmentation into cells and, at the receiver, correctly reassemble the cells.The requirement for correct reassembly is generally necessary if the adaptation layer is to fulfil its nominal role of adapting ATM cells to the variable-length PDUs of conventional protocol stacks, but we question the requirement for accurate reassembly more closely in Section 5.1. The issue is a hardware versus performance tradeoff predicated on the desirable complexity of an ATM interface which performs or helps with adaptation in hardware.
A suitable data adaptation layer is the ATM adaptation layer type 5, being standardised within CCITT [2]. AAL-5 uses the bit to indicate last cell of a block, where the last cell includes a length field (or cell count) and a 32-bit CRC over all the cells forming the block. The length field is used by the adaptation layer to determine that the correct number of cells have been received. If, necessary, a further length field, indicating the valid data section of the PDU, etc., can be placed elsewhere in the PDU by higher level transport or RPC protocols.
The CRC detects the usual bit and burst errors, as well as most cell
reordering while the length is required to detect certain cell drop
outs where the content and position of the cell contrive to provide no
contribution to the CRC. The CRC polynomial used in an ATM adaptation
layer needs to be chosen with care, since if the CRC of 
is zero, swapping two cells distance 
apart will not be detected.
The polynomial selected for AAL type 5 (the standard CCITT CRC32) will
detect such pair swaps up to the maximum size of the AAL5 PDU, 64K
bytes.
We have measured the rate at which a workstation can calculate a CRC in software. Calculation of a 32 bit CRC on blocks of data on a 25MHz DECStation 5000/200 can be done at only 13 Mbit/s. This is using an 8 bit at a time method and two 256 by 32 look up tables. (Using 16 bits at a time requires two 64K by 32 look up tables which have undesirable effects on cache performance.) This suggests that assistance with the calculation of AAL CRCs is required in the interface.