Reviews for paper Glacier: A Query-to-Hardware Compiler, submitted to SIGMOD 2010 (Demo Track).
Overall rating: accept
yes
Accept
Weak Accept
Weak Accept
yes
Weak Accept
Accept
yes
Glacier is an impressive system for compiling streaming queries directly into FPGA-based hardware. Even though the details of the system has already been published, demonstration at sigmod would provide attendees a good opportunity to see all the steps of hardware generation in action.
The demonstration would be more impressive if it included a discussion of the Glacier limitations in the pre sense of limited gate budget. For example Xilinx FPGA may not have enough gates to generate TCAM large enough to hold all the unique groups in grouping operators. Large query sets may go above the gate limit and require non-trivial optimization to figure out which predicates and operators and more important to be pushed into hardware.
yes
Accept
Accept
Accept
yes
Weak Accept
Accept
no
The demo presents interesting, novel technology in an easily accessible way, and much thought went into making it compatible with a conference demo slot. It is based on a arecent VLDB paper, and includes a general demo script and interactivity opportunities. It should be a fun demo.
SOme sort of comparison with software-based approaches would be nice...
yes
Weak Accept
Accept
Accept
yes
Accept
Weak Accept
yes
The proposal is to demonstrate Glacier (pvldb 2009), which is a query-to-FPGA compiler for streaming queries. The demo will illustrate the details of this compilation process. This would be a good educational experience for database audience to see FPGA in action.
One of the major motivations of using FPGA in Glacier is to deal with high network package rate. The observation is that the network stack on a commodity machine is not good enough to handle such high rate.
However, recent work in the networking community (RouteBricks, SOSP'09) shows that commodity machines can be used to implement a 35Gbps network router. If packet switching routers can be implemented with commodity machines, one would imagine data stream processing can be supported too. Therefore, the above motivation needs re-evaluation.