Moving ProcessingTowards Data
Vijayshankar Raman, Inderpal Narang, Prashant Pandey · 2008
Storagearchitecture includes moreand more prrocessing powerforincreasing requirement ofreliability, man- agibility andscalability. Forexample, anIBM storage server is equipped with4 or8state-of-the-art processors andgigabytes ofmemories. Thistrendenables analyzing datalocally inside a storage server. Processing datalocally isappealing underthefollowing cir- cumstances: 1)hugereduction ofdataflowing tothehost, 2) reduction ofCPUconsumption onhost. Accordingly, thebenefits are1)less datatraffic through 10channel tothehost, 2)better utilization ofhostbufferpool, and3)enabling moreworkload on thehost. Onecrucial taskistounderstand howDBMS canbenefit from suchhardware. Thatistoidentify whichdatabase operations are beneficial tobeoffloaded given aqueryworkload inaparticular setting. Forcertain operations, weestablish value proposition via various approaches andshowtheanalytical andexperimental results. Inparticular, starjoin queries arecommonlyusedin business warehouses. Wepropose tooffload aportion ofastarjoin queryfromhosttothePOWER5 P processors ona storage server, whichdramatically reduces theamountofchannel 10 andhostCPUconsumption. Moreover, thequeryelapsed timeis improved viatheexploitation ofthestate-of-the-art Pprocessors onastorage server.