Testing Other Internal Switch Modules Clause Samples

Testing Other Internal Switch Modules. A similar process is followed to generate deterministic test patterns for the port arbiters, the LBDR modules and the crossbar. Also the implementation of their TPGs is identical, and so are the optimization techniques. Again, the most relevant practical implementation issue concerns the communication of test patterns or responses across the switch-to-switch links for the crossbar and LBDR module. The crossbar outputs 34 bits in response to a test vector: 32 data bits, 1 valid bit and 1 stall bit. The communication channel can only carry 32 bits (the valid bit of the channel needs to be permanently set to 1 during test vector transmission, while the stall signal travels in the opposite direction). The two remaining crossbar signals (valid and stall) which do not fit into the link can be either transmitted by means of additional lines used only during testing, or alternatively checked by local comparators, similarly to what has been done for the communication channel. We took the latter approach, and the results in section 2.4.1 again confirm the marginal coverage reduction on TPG faults. Fault coverage of the crossbar is not affected at all by this choice. Unlike other modules, test vectors for the LBDR modules should be transmitted across the link, and they take 31 lines (the primary inputs of the LBDR module). So, they perfectly match with the current flit width, provided the number of network destinations does not exceed 64. From there on, the test vector width starts growing logarithmically with the number of destinations, and additional lines may be required on the link. In contrast, the use of a larger flit width in the network (e.g., 64 bits) would automatically solve the problem. In that case, the test patterns of the LBDR block and the test responses of the arbiter could even be communicated at the same time over the link. Also, since LBDR module and arbiters have only few outputs, their response checking could be performed at the same time on the available tree of comparators, thus cutting down on the test application time (see section 2.4.1).