The meaning of UNC_IMC_ISOC_FULL.READ.CH0 on Xeon E5520

[emy_lv]

I runned 4 tasks on one of the two socket on E5520. These tasks are independent and they are both cache-missintensive and bandwidth intensive. There is very trivial file operation which may be ignored. The other socket is idle during the running of that task. However, Iobserved avery high number of UNC_IMC_ISOC_FULL.READ.CH0 and UNC_IMC_ISOC_FULL.READ.CH1 onboth sockets with pfmon3.9 . The number on both sockets are almost the same as if they are from the same counter. Could someone give me some explaination on this event?

BTW, I read the following lines in the performance manual of Intel
"When first programming the queue occupancy events, a bus lock must be issued to correctly initialize the occupancy. This is the same requirement encountered with the GQ and QHL occupancy events." Need I make the bus lockwith pfmon and how?

[Thomas_W_Intel]

UNC_IMC_ISOC_FULL.READ.CH0 counts cycles all the entries in the DRAM channel 0 high priority queue are occupied with isochronous read requests.Are you maybe accessing memory that is attached to the other socket? The OS usually tries to allocate local memory, but if memory interleaving is enabled or your run out of local memory, then remote memory is used.