tcp-ip@ucbvax.ARPA (07/19/85)
From: Jerry Morence <morence@Almsa-2> Richard: Regarding your message dated Sun 30 Jun 85 19:18:05 PDT, subj: TCP/IP on Hyperchannel. In ref. msg., you stated you had problems with message collision which resulted in manual resets of the hyperchannel adapters. In three years of operation at six different sites, we have never experienced this problem. Our traffic workload at each of these sites is measured in gigabytes on a daily basis. We only use A220/A222 adapters with a dead-man-timer of 500 milliseconds. Our hosts must be able to retrieve all messages in this time frame or the message could be lost. The A220/A222 adapters recognize when a buffer at a destination adapter is full and report back to our software that the message can't be delivered; our software will requeue the message for retry within one second. This will continue until the destination buffer is freed either by the destination host retrieving the message or by the dead-man-timer causing an automatic release (and potential loss). Our software recognizes when the dead-man-timer has expired and requests a retransmission from the originator; the information for this request is contained within the message proper for the lost message which is held by the A220/A222 adapter in a separate buffer (there are eight of these type buffers). In the case where our software on the send side may do a retry at the same time our software on the receive side requests a retransmission, our software detects this situation and prevents sending/receiving a duplicate. If an application crashes, our software recognizes this and takes action to quiesce all connections to the crashing application without impacting other connections in such a manner that facilitates recovery and does not demand termination of the connected applications. In all these years of production, we have never needed to reset the adapters, we have never lost nor duplicated data, and we are achieving a minimum of 30 megabits per second of real data throughput. In the near future, we will be experimenting with an A440 adapter, connecting a VAX 780 to the Hyperchannel. Maybe we will then experience the same problems you have been reporting. Regards, Jerry