MAMEWorld Forums - The Loony Bin - Re: New toy (geek pr0n)

> Interesting. Though I'm kinda curious why one would bother these days with Wifi an all.....

Wifi? Hahaha! That NIC is at least an order of magnitude faster than any wifi. With wifi all devices on the network are sharing the same aggregate bandwidth, whereas with copper/fibre each link has full dedicated bandwidth. You don’t need to worry about identifying and authenticating with the network when it’s taken care of by what you plug in where. Wifi suffers from wireless propagation and interference issues. Also wifi latency is bad because of the channel sounding and training sequences needed to try pushing high bandwidths over radio.

High-speed networking has a number of applications:

High-speed data acquisition, where sensors produce data at very high rates and you have to get it to storage/processing (e.g. radio astronomy, particle accelerators)
Storage centralisation/virtualisation, where you have all your disks managed by storage appliances and access it over the network (e.g. data centre or home SAN)
Distributed computing where you need rapid synchronisation between nodes – the interconnect is usually the bottleneck in these cases
Latency-critical applications where you need to respond to an input as quickly as possible (e.g. options market making, spread trading, statistical arbitrage)

That Chelsio NIC accelerates a number of common use cases. It does iSCSI and FCoE allowing you to use networked storage without the OS having to worry about the network part of it.

The Chelsio TCP offload engine takes care of window management, stream reconstruction, checksums and retransmissions so the OS doesn't have to do this. That alone saves significant CPU resources, memory bandwidth and cache thrashing. On top of that it also allows the final data to be DMA’d straight into the application’s address space, avoiding slow transitions between kernel mode and user mode and minimising copying. This improves wire-to-application latency enormously.

You can also go the other way and map a ring buffer to receive raw frames with no preprocessing applied. You can disable the NIC’s interrupts, run hot polling the buffer, and never sleep or get scheduled out by the kernel (so you never end up in a situation where data is available but you need to wait to be scheduled to process it).

So as I said, it’s a specialised piece of kit and not exciting for most people, but it does what it does very well. You have to be a geek to appreciate it.