3-in-1 Kernel - 2023.2 English

The 3-in-1 GQE kernel is a compound of multiple previous released GQE kernels, containing a large number of post-bitstream programmable primitives. It can execute not only hash-join/hash-bloomfilter/hash-partition individually but also hash-based bloom filter build + partition or bloom filter probe + partition as a combination to minimize the intermdediate data transfer. To instantiate the 3-in-1 in the resource-limited U50 Alveo for obtaining the best benefit-cost ratio, we refactored every processing unit (PU) as well as the output data paths and retired the bypass design in the current kernel, now it enpowers the Q5-simplified query can be done with a 4x performance improvement without any cost increasing on device comparing to the previous separated GQE kernels.

The internal structure of 3-in-1 GQE is illustrated in the figure above. Besides those challenging parts we have accomplished in previous separated GQE kernels, it is important to emphasize the necessity of the hardware structure we implemented in 3-in-1 GQE for reusing the 2 AXI-Master ports and 3 internal huge URAM buffers in each PU, as we cannot tolerate even a little of resource wastes on the device to make the gqeJoin/gqeBloomfilter/gqePart all on a single U50 comes ture.

The reason why we need this reusable hardware structure is because of the limitation on the resources of U50, and the different precedence of URAM and HBM in different flows. It can be described in the figure above on the left side of the vertical red line.

JOIN build/probe flow: needs to save the total number of an unique key in the URAM first, and then save the key/payload pairs in the HBM.

Bloom filter build/probe flow: needs HBM to access its corresponding hash-table.

PART: needs URAM to buffer the post-partitioned key/payload pairs to get a reasonable throughput when flushing the partitioned bucket out.

While performing the bloom filter probe operation + PART flow, we don’t know when the key/payload pairs in a specific bucket is enough to be flushed out. To avoid implementing a duplicated URAM buffer after the partition module, we have to arrange the bloom filter operation before the partition so that we can fully utilize the original URAM in partition module to collect the partitioned rows. Thus we need this reusable hardware structure as illustrated on the right of the vertical red line in the figure above, where the data goes along with the different paths (marked with red) under different configurations.

The hardware resource utilization of 3-in-1 kernel is shown in the table below (post-placement). Total part contains not only the listed sub-modules resource utilization but the interconnect streams to serve for the dataflow design in 3-in-1 GQE kernel.

Module	LUT	LUT as memory	Register	BRAM36	URAM	DSP
AXI adapters	34306	9256	75782	159	0	0
Scan	9581	900	9581	1	0	0
Filter	20480	3936	11850	0	0	0
crossbar 4-to-8	9044	15110	4696	0	0	0
PU x 8	285419	47728	257620	300	264	72
adapt sizer x 8	9597	0	8476	0	0	0
collect 8-to-1	7147	0	8543	0	0	0
Write	19205	1612	22488	1	0	3
Total	424564	81037	434954	569	264	75