Volume based on/off cache prefetch

Erste Group is building up the new HW and SW infrastructure for corporate DWH system
For hot data they are using 8 node SVC (model DH8) cluster and FlashSystem 840 as the SVC back-end Tier0 storage. For the cold data (Tier 1) they are using the same SVC cluster with V7000 Gen2 as back-end storage

Customer performed following tests on the new infrastructure before movement into the production :

IBM Power 8 servers and SVC virtualized storage (Tier1):
LOAD:100% random reads, block size 1MB,
3 threads/LUN
LPAR: 6x300GB no compressed LUNs,
8x 16Gbps active paths to hdisk
SVC: 3 IOgroups,
V7000: 72 pcs of 4TB 7,2k 3,5" NLSAS HDD in the six RAID6 arrays, (plus two spare disks)

There was measured huge discrepancy between front-end and back-end throughput on the SVC level
Backend mdisk throughput was 1200MB/s,
Frontend vdisk throughput was only 520MB/s

On the Host level were measured the similar numbers like in the SVC front-end:
LPAR throughput : 523 MB/s (LPAR ~ SVC vdisks)
LPAR number of IOPS : 2100 (LPAR ~ SVC vdisks)
backend throughput : 1200 MB/s (SVC mdisks ~ V7000 vdisks)
backend number of IOPS : 5000 (SVC mdisks ~ V7000 vdisks)

On the V7000 level was measured values:
V7000 drive throughput: 2500 MB/S (V7000 mdisks)
V7000 drive IOPS : 10000 (V7000 mdisks)

Test results:
On the disk drive level we achieved throughput 2500MB/s. This throughput would be acceptable by the customer But this value is decreased to the 500MB/s by V7000 and SVC.
After disabling of the R/W cache at storage system level performance was increased.
Values achieved with SVC and Storwize R/W cache disabled: throughput was between 2700 MBps and 3000 MBps at the host level. Those are values acceptable by the customer