Traditionally, Content Delivery Networks (CDNs) have focused on achieving competitiveness by driving system performance up and driving hardware costs down. From a hardware perspective, the focus has always been on three specific attributes: transfer rates (measured in packets per second, or PPS); power consumption of the total solution (less is better), and total hardware capital expense (CapEx). CDNs mechanized this by buying servers from original design manufacturers (ODMs) such as Hyve Solutions, Ciara, and Quanta Cloud Technology (QCT), while providing a list of qualified vendors to the ODMs. It was then up to the ODMs to select the right components for their offering, and to get the lowest price that they could negotiate. The ODM with the best competitive bid that also met the CDN’s power requirements would win the bulk of the orders for that time period (usually between one quarter and one year). While this approach produced what was arguably “best in class” servers from a price and performance standpoint, they still do not achieve the costs that were achieved by hyperscale cloud companies.
To remain competitive, CDNs need to look to technologies that can give them orders of magnitude improvement in performance. One of the most promising of these is Computational Storage, where multicore processors are embedded within modern NVMe solid-state drives (SSDs). Beyond the simple fact that storage is a requirement to keep up with the data, so buying SSDs happens regardless, adding compute via Computational storage brings two significant advantages to CDNs.
The first of these is that computational storage devices allow processing capacity to be scaled linearly with storage capacity without adding more expensive servers. This allows CDNs to either increase the total transaction capacity of their existing server farms, or to reduce the footprint of their points of presence (PoPs). Since most PoPs are in colocation facilities, this can significantly reduce the real estate, power, and cooling costs for the CDN.
The other advantage that Computational Storage brings to CDNs is significantly reducing data movement within their PoPs. Most of the data movement within a PoP is not from the CDN’s servers to customers outside the PoPs – it is from storage devices to the servers. This movement is necessary today to perform data encryption, which is necessary for the delivery of streaming media with digital rights management (DRM) limitations. It is also necessary today when performing searches. Most of this processing can be moved to computational storage devices, which can significantly speed up these transactions. We will examine the magnitude of this performance improvement in our next blog in this series.