Anyone have TCP Offload Engine (TOE) experience?

[LAVA 1.0 Content]

I've been slowly adding higher and higher data rates to my multichannel UDP receiver and have pegged the CPU (95-100% while streaming). The effective rate is 3 channels * 4800 messages/sec * 512 bytes/message or ~ 59 Mb/s. I'm looking to free up some cycles for additional features. I recalled reading some time ago about smart NICs that offload some of the TCP processing from the CPU, so I started Googling.

The proper term seems to be TCP Offload Engine (TOE). No fetish jokes or links to "inapropriate for work" sites please

Anyone have experience with TOE? So far I've only found PCI-X NICs and references to 64 bit Windows; I'm more interested in a PCI / XP Pro solution. I can't change out my processor or backplane at this time; I'm using a P4 3.0 GHz, 2 GB RAM and in a rackmount PC with passive backplane (PCI only).

[Matt W]

QUOTE(LV Punk @ Sep 6 2007, 05:26 AM)

I've been slowly adding higher and higher data rates to my multichannel UDP receiver and have pegged the CPU (95-100% while streaming). The effective rate is 3 channels * 4800 messages/sec * 512 bytes/message or ~ 59 Mb/s. I'm looking to free up some cycles for additional features. I recalled reading some time ago about smart NICs that offload some of the TCP processing from the CPU, so I started Googling.

The proper term seems to be TCP Offload Engine (TOE). No fetish jokes or links to "inapropriate for work" sites please

Anyone have experience with TOE? So far I've only found PCI-X NICs and references to 64 bit Windows; I'm more interested in a PCI / XP Pro solution. I can't change out my processor or backplane at this time; I'm using a P4 3.0 GHz, 2 GB RAM and in a rackmount PC with passive backplane (PCI only).

In my mind the best option would be to upgrade the CPU, to a Core 2 based one (if possible), but since that doesn't seem to be an option now.

I've never used one, and have no idea how well it would work (it's built for low latency gaming, so I don't know what it'll do in your use case), but a http://www.killernic.com/ goes beyond TOE. It basically runs a linux os on the network card that can offload a lot of the network processing (it could offload the firewall for instance in addition to the tcp stack). Also It's PCI and XP compatable.

[LAVA 1.0 Content]

Thanks! This statement is interesting; From http://www.killernic.com/support/faq.aspx#Q49

Q: Who else would benefit from using this product?

A: Among other things, Killer is a UDP (User Datagram Protocol) accelerator, so other applications that are UDP intensive could also see benefits. In addition, programmers, Linux developers, and other enthusiasts will love FNA because they will be able to write or download Flexible Network Applications that run on Killer's NPU. Flexible Network Applications could be VOIP apps, peer to peer file sharing utilities...the list is endless. FNA is the infinite potential of Killer.

I'm not sure my customer (Uncle Sam) would commit to using this; but it looks good. Most TOE stuff requires a high-end server config (that I don't have).

The dual-core (or quad, what the heck!) has been in my mind; I might be able to get my supplier to let me eval and check the performance difference.

[Matt W]

QUOTE(LV Punk @ Sep 7 2007, 05:04 AM)

Thanks! This statement is interesting; From http://www.killernic.com/support/faq.aspx#Q49

Q: Who else would benefit from using this product?

A: Among other things, Killer is a UDP (User Datagram Protocol) accelerator, so other applications that are UDP intensive could also see benefits. In addition, programmers, Linux developers, and other enthusiasts will love FNA because they will be able to write or download Flexible Network Applications that run on Killer's NPU. Flexible Network Applications could be VOIP apps, peer to peer file sharing utilities...the list is endless. FNA is the infinite potential of Killer.

I'm not sure my customer (Uncle Sam) would commit to using this; but it looks good. Most TOE stuff requires a high-end server config (that I don't have).

The dual-core (or quad, what the heck!) has been in my mind; I might be able to get my supplier to let me eval and check the performance difference.

My main concern with it would be reliability (I heard the early drivers were flaky, I'm not sure if that's still the case,but only way to be sure is to test it). Also the standard (32 bit, 33 MHz) PCI bus only has 133 MB/s of bandwidth (gigabyte has a theoretical peak of 125 MB/s), so if you have anything other than a Gigabyte NIC on that bus you could run out of bandwidth. In older mother boards the hard drive controller would typically hang off the PCI bus, which would be limiting if that's the case on the current motherboard.

Even a Slow 1.8 GHz Core 2 should be faster than a 3.0 GHz p4 in the majority of cases (including running with only one core). Let alone a 2.4 GHz quad core (the cpu only costs ~$280). But I doubt your motherboard supports it (even though it may have the right socket).

You can get an idea of the speed difference here.

http://www23.tomshardware.com/cpu.html