Host Packet Buffer (HPB)
In order to support multi-core CPUs and multithreaded host applications, ANIC adapters utilize a flexible host packet buffer (HPB) technique. Host memory is segmented into a number of fixed size blocks. The block size is configurable but is typically 2MB or 4MB each. A collection of these host memory blocks is then dynamically pooled together to form a host packet buffer. A specific application thread (often tied to a CPU core) is then explicitly assigned or linked to a given HPB and will only process data that is transferred into its own HPB. Up to 64 independent HPBs can be created (per ANIC adapter) and in turn assigned to up to 64 host application threads.
The memory blocks (typically 2MB or 4MB in size) assigned to a host packet buffer (HPB) do not have to be contiguous. In other words, each HPB is composed of blocks of host memory that are randomly spread out in various areas of physical memory. In addition, memory blocks are temporarily assigned to a given HPB by the ANIC adapter and once an application thread has finished processing all the data from a given memory block, that block can be assigned to a different HPB.
An ANIC adapter is configured to intelligently steer packets in to specific host packet buffers (HPB). The benefit of packet steering is that each thread in a multithreaded application (often utilizing multiple CPU cores) can process packets from its own HPB. In this way a security or networking application can take advantage of parallel processing of data thus achieving higher levels of speed and efficiency.
There are three different ways to steer packets into a HPB:
- ANIC adapter is configured to use its own internal algorithms to evenly and efficiently distribute or load balance packets across a specified number (from 1 to 64) of HPBs. This is done to ensure that no processing thread is overwhelmed with data while others are starved.
- Based upon the results of packet filtering, packets can be steered to specific HPBs. For example, packets that match a specific packet filter rule might all be steered to the same HPB for processing.
- Based upon flow classification, packets are steered to specific HPBs. In other words, specific flows are identified and explicitly steered to a specific HPB for processing.
Packet traffic is typically transferred across the PCIe bus (DMA) for consumption by the host application. However there may be circumstances under which select traffic must be locally redirected or retransmitted out of one of the ANIC network ports. Packet filtering or flow classification can be used to identify which specific packets or flows must be redirected out a given port.