Server virtualization, which concentrates multiple applications and the infrastructure functionality of network processing in a general-purpose server environment, is spreading beyond datacentre to the field of edge computing, including wireless base stations and Mobile Edge Computing (MEC).
The technology is also being used with applications like traffic management, and in entertainment, where it can be used to deliver viewers content like sporting events. As data volumes increase and systems become increasingly complex, however, the CPU resources required for packet processing in a virtual network increase, which in turn reduces the number of applications that can run on a single server, and leads to lower server aggregation rates.
Fujitsu Laboratories has developed technology to accelerate packet address control, which had been a performance bottleneck, while also offloading the processing of router functions in the virtual network to field programmable gate arrays (FPGAs). Fujitsu has been able to speed up packet processing performance eighteen-fold compared with existing virtual routers, while reducing the use of CPU resources to about one thirteenth that of existing technology.
Previously, high-speed processing was conducted by subjecting all packets to the same processing, without changing the order of the packets. If the order of the packets were to change in the input or the output of the virtual router processing, that might lead to applications resending packets, which would lower quality and increase the burden on the system.
With this new technology, packet sequence information is stored separately from the packets in the pipeline. Consequently, in address search processing spread over multiple stages, when the results of an earlier search render subsequent searches unnecessary and a packet bypasses those searches, it can be rapidly recombined with the other packets in the pipeline in the proper order based on the packet pipeline sequence information. Combined with the pipeline duplication, this technique reduces memory access while maintaining high pipeline processing speeds. These technologies improve packet processing performance.
The technology is able to automatically switch between the high-speed, low-capacity memory within the FPGAs and the high-capacity external memory, without pausing address search processing. In address search processing, multiple tables are searched to determine the address. As the number of connections increases, the amount of memory used by the lookup tables increases and the free space in the internal memory becomes insufficient. When this happens, this technology can synchronize lookup tables with relatively low access frequency, relative to their size, with external memory in the background, automatically switching to lookup tables in external memory without stopping search processing. This enables stable packet processing performance, limiting access to external memory, even in cases such as large-scale systems, where there are many communication destinations, and large lookup tables become essential.
An evaluation of the effects of this offloading on a general-purpose server was conducted using this technology, which was deployed on Intel Stratix 10 MX FPGA featuring high-speed HBM2 memory using Tungsten Fabric, an open-source virtual router.
Two servers connected with 100Gbps ethernet ran four virtual machines each, and communications were conducted between each virtual machine to test the performance of the virtual routers. The results showed that packet processing performance, which was 13.8Mpps using existing methods, was increased to 250Mpps, or about an 18-fold increase. In addition, the number of CPU cores used was reduced from 13 cores to 1 core.
Using this technology, it is now possible to increase application server density. This will support the digital transformation of customers from the foundational level, improving server utilization efficiency in infrastructure business areas that use large volumes of data in the 5G era, particularly in base stations and MEC for carrier businesses, such as, for example, making it possible to operate a video distribution service for a stadium with fewer servers.
