Hi, We are software developers, part of Marvell's cellular platform infrastructure team.
Our team has been working on a project named "fastpath" for speeding up IP forwarding in embedded systems. The initial version (fastpath v1) has already been successfully deployed in our latest pxa1801 (cellular modem) based products. We are in the final stages of fastpath v2 development, which is completely hardware independent and requires minimal changes in the generic networking code (the project consists of a kernel module and a single kernel patch); despite being hardware independent, fastpath v2 already achieved the same level of performance (as fastpath v1) and even increased stability. Our development platform is running openwrt Barrier Breaker (r43694), so naturally we chose to suggest this to the openwrt development community first. You can find a brief description of our fastpath solution below. We are anxious to hear your thoughts/comments and will gladly share the code. Best Regards, Ram Marzin r...@marvell.com Tomer Eliyahu tom...@marvell.com Fastpath in a nutshell ---------------------- The basic concept of fastpath is to optimize the data-plane while keeping the control-plane in the generic networking stack. This is a known concept in the industry which is commonly used in embedded systems [1], but so far we couldn't find any open source implementation for it. Fast path implements an optimized data-plane, which replaces the generic data-plane forwarding code for selected connections. The data-plane implementation includes a straight forward optimized packet processing engine which handles all the required packet manipulation for IP forwarding, such as decrement ttl/hop count, checksum adjustment, MAC header encapsulation and "dummy NAT" (TCP/UDP traffic which does not carry any L3/L4 information in the packet payload). As noted above, the control-plane is handled by the generic networking stack, with the only exception of learning new connections and marking the valid ones as fastpath - some connections can't participate in fastpath, such as any "non-dummy NAT" connections (e.g. FTP control port), local traffic, and any protocol which is not supported (e.g. IPv6 extensions, IPSec, IPv4 fragmanted packets, etc.). Needless to say that ALL non-fastpath connections / protocols will work as is, i.e. they simply won't go through fastpath. As a rule of thumb, it is safe to assume that in most of the cases, 90% of the data will go through fastpath. In our experiments on pxa1801, fastpath alone *almost doubled* the performance (both Throughput and MIPS consumption) for TCP/UDP IPv4/IPv6 forwarding. References [1] http://www.embedded.com/design/operating-systems/4403058/Accelerating-network-packet-processing-in-Linux _______________________________________________ openwrt-devel mailing list openwrt-devel@lists.openwrt.org https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-devel
