The following settings should only be modified after research and understanding of link and replication characteristics and under the guidance of Compellent personnel.
This is the number of sectors to transfer in a single IO request. This is the maximum size of an IO request. A bigger number causes less "chopping", conversely, a smaller number causes more chopping to be done (in other words, smaller chucks of data will be sent on the wire). This number has a relationship to the target side write cache. The write cache on the target side is current set to 256 sectors. Therefore, if you set this value to more than 256 (the maximum cacheable block is 128 sectors), you will force a bypass of the far side write cache. For a T1 link, 400 sectors/sec will saturate the wire.
The outstanding IOs per destination. This is related to the execution throttle set in the HBA. For iSCSI, the iSCSI protocol limits the number of requests to 64 outstanding, so setting this value to larger than 64 does not affect iSCSI transport. This does however affect FC transmissions. Note that the size is just short of 128 to allow room for Scsi initiator and IPC requests.
>The total amount of data in the pipe at any given time. If you have 16k outstanding sectors, this will imply 8M bytes of memory is used on the source controller. If (Max sectors per IO * Max outstanding IOs) is greater than Max outstanding sectors, then outstanding IOs are limited.
Maximum number of IOs outstanding for each destination. Scsi initiator limits this to 16, so a value of 18 allows for some small amount to be outstanding at any given time.
Maximum number of sectors outstanding for each destination. This parameter is related to the “Max IOs per destination” since both impact the how much total IO is sent towards the destination.
This defines the number of IOs to satisfy at each pass through the system queue. This is relevant on a per destination basis. For example, if there are 3 volumes to a single destination, and they each have IO outstanding (6, 4, and 4 respectively), then, on the first pass through, if the Max IOs per queue pass is 8, then all 6 IOs on the first volume will get satisfied and 2 IOs from the 2nd volume will get satisfied. On the next pass through, if 3 IOs have completed, then 3 IOs from the 3rd volume will be accepted and so on. By using a small number, we ensure all volumes to a destination get serviced often.
If this number is too large, on the 1st pass, the first volume will get serviced while the other volumes starve. On the 2nd pass the 2nd volume would be serviced but the 1st and 3rd volumes would starve.
This limits the amount of data sent towards a destination in a single pass through the system queues. It is set equal to the Max sectors per destination. Smaller values for the Queue Pass parameters increases fairness at the expense of CPU utilization