Takeaways

1. Some words

   linear throughput scalability. on-the-fly data transformation

2. How can we integrate ML into a column-store DBMS without disrupting either DBMS efficiency or ML quality and performance?

System

It integrates ML into a column-store DBMS without disrupting either DBMS efficiency or ML quality and performance.

1. The paper propose pSCD to achieve the cache-efficient training on column store
   1. For a batch of data, read each column and then perform training requires Batch*Feature cache size.
   2. Partitioned SCD to achieve cache-efficient training: Coordinate-descent based algorithms enable a way of accessing the samples one feature at a time, which natively corresponds to column-wise access.
2. The paper uses FPGA to do the data-preprocessing and ML computations.

---

• ← Previous Post
• Next Post →