Oct 30 2018
In our last blog, we discussed some of the ways that computational storage can improve transactional capacity and reduce costs for Content Delivery Networks (CDNs) within their standard storage growth. In this blog, we will examine one example of how NGD Systems computational storage solid-state drives (SSDs) increased performance for an application known as Facebook Artificial Intelligence Similarity Search, or FAISS. Image similarity searches are the next generation of search capability for websites – it allows users to find images similar to one they already have. In this case, due to the proximity of the US Holiday Halloween, we can use the idea of searching for a pumpkin cutting template. The uses of FAISS extend beyond searching for people on Facebook (its name notwithstanding ). The “magic” of FAISS is that it significantly improves the matches returned through its use of AI, while significantly reducing the memory footprint required for the searches. Again, finding your favorite pumpkin cut out faster!
To test the effectiveness of computational storage for accelerating FAISS, NGD Systems utilized a COTS hardware solution, but simply replaced the SSDs with Computational Storage devices.
The impact of computational storage on FAISS is shown in the diagram below. As can be seen, the (H+1, H+2 … H+16) configuration with sixteen Computational SSDs provided nearly constant response times of half a second. In contrast, the configuration that didn’t utilize computational storage saw its response times grow exponentially, with the maximum response time of 42 seconds. Imagine what that means to all the people searching for that pumpkin image. 42 sec to hundreds, thousands, or millions of people adds up fast!!
The greatest driver of this increase in response time was the need to move data from the SSDs to CPU memory over the PCI Express (PCIe) bus. Computational storage significantly reduced this data path by eliminating the need to move data from storage, index the data, and load the data into the database. For a CDN, this performance improvement translates directly into significantly increased transactional capability (a roughly 80X improvement), or a significantly reduced footprint in their PoPs for image similarity search transactions.
Another way to look at this, if I had a server with 4 Computational Storage drives, I would have to add another server for every four drives to ‘maintain’ response time. While with NGD Storage devices we provide the same response regardless. Therefore, saving the need to buy 4 full server platforms, an excessive CapEx cost to the CDN.
To find out more on how NGD Systems can improve the performance of your CDN infrastructure, please visit our website.