Would that not push the bottleneck to the bus on the compute server? I haven't gone through the specs to see whether that it would saturate out otherwise, but that would be my guess.
It's not about the particular HW config. The point of the machine is that it moves a portion of querying to the disk, so the disk is shipping back matching tuples as part of the parallel query, rather than blocks. There are a couple of CPUs, 4 cores each on each disk.
The great problem in high-scale data warehouses is in shipping blocks across the interconnect. This reduces that.
If somebody builds a low-priced 8-core shared memory multiprocessor, the Cell guys are toast
how does sharing cores address the shared cache coherence problem?
Slashdot Mirror
Would that not push the bottleneck to the bus on the compute server? I haven't gone through the specs to see whether that it would saturate out otherwise, but that would be my guess.
The pipe is the bottleneck, not the compute server. I assume they do this because more servers allows more pipes between storage and dbms.
It's not about the particular HW config. The point of the machine is that it moves a portion of querying to the disk, so the disk is shipping back matching tuples as part of the parallel query, rather than blocks. There are a couple of CPUs, 4 cores each on each disk. The great problem in high-scale data warehouses is in shipping blocks across the interconnect. This reduces that.
If somebody builds a low-priced 8-core shared memory multiprocessor, the Cell guys are toast
how does sharing cores address the shared cache coherence problem?