Slashdot Mirror
'Killer' Network Card Actually Reduces Latency
fatduck writes "HardOCP has published a review of the KillerNIC network card from Bigfoot Networks. The piece examines benchmarks of the product in online gaming and a number of user experiences. The product features a 'Network Processing Unit' or NPU, among other acronyms, which promise to drastically reduce latency in online games. Too good to be true? The card also sports a hefty price tag of $250." From the article: "The Killer NIC does exactly what it is advertised to do. It will lower your pings and very likely give you marginally better framerates in real world gaming scenarios. The Killer NIC is not for everyone as it is extremely expensive in this day and age of "free" onboard NICs. There are very likely other upgrades you can make to your computer for the same investment that will give you more in return. Some gamers will see a benefit while others do not. Hardcore deathmatchers are likely to feel the Killer NIC advantages while the middle-of-the road player will not be fine tuned enough to benefit from the experience. Certainly though, the hardcore online gamer is exactly who this product is targeted at."
I have a friend who works at a game studio who had their IT guy perform several tests to see if it did anything... chat log:
> killer NIC. bad.
> file transfers = 1/4 of the speed of a normal NIC
> the drivers are fucking TERRIBLE to install/uninstall/update, you have to reboot. then it'll let you do whate you need to do. then reboot AGAIN...
> and when it does start working, there is literally no difference in either framerate or ping, even on the games they say it specifically improves
Anandtech has a much better review here: Linky: http://www.anandtech.com/cpuchipsets/showdoc.aspx? i=2865
It seems a like a lot of what this NIC claims to do (offloading TCP work such as checksumming and segmentation) can be done by an Intel NIC on Windows and Linux (2.6.19 and up at least). So go get a gig Intel NIC for the $35 and invest the rest in a new mouse pad or something that might actually impact your gaming a bit more than that silly nic. http://shopping.yahoo.com/p:Intel%20PRO%2F1000%20P T%20Desktop%20Adapter:1993037566;_ylt=Apu.Yd0SJ9f3 4z.56feYYLIbFt0A;_ylu=X3oDMTBic2hxMGNhBGx0AzQEc2Vj A3Ny?clink=dmps/intel_pro.2f.1000_pt/ctx=mid:5,pid :1993037566,pdid:5,pos:2,spc:14489115,date:2006120 9,srch:kw,x:
Tim Smith - Ramblings from Nerd Land
A better design would be to have networked data in well-defined regions which the card can DMA directly out of and into. The PCI bus can handle a 4K transfer as an atomic operation on a single channel, so a 4 channel PCI card can simultaneously send and receive streams of jumbo packets without requiring any CPU intervention. The driver would merely need to be passed two lists of physical page pointers - one for inputs, one for outputs - for each of the open connections and to pass back signals from the board that an entire packet and/or message had been uploaded for a given connection.
(Jumbo packets can go up to 8K, and you can do 8K over two PCI channels, so that's a good unit to be working with.)
The next improvement that can be made to such a system is to improve the buffering. The network will be slower than the computer, almost always, so being able to queue up multiple packets for sending is a Good Thing. Filtering out packets that have not been sent but are no longer worth sending can be done entirely in parallel and does not require anything extra at the end of the pipeline. Not all packets are of equal value, though. You want to deliver the information in the order that will give the best possible benefit - which may not be the order in which the program generates the traffic. Hierarchical Fair Service Curve, Class Based Queueing, and a bunch of other similar techniques, have been developed to fix exactly that sort of problem.
If the ISPs would stop being so bloody stupid, you could also enable protocols such as your basic multicasting (for your UDP stuff) and Scalable Reliable Multicast (for the stuff that needs to reliably get through). That hacks, slashes, butchers and roasts (with just a hint of parsley) problems associated with sending identical state information to multiple end-points.
Bear in mind that Myrinet, Dolphinics, and a bunch of other vendors, use essentially the above mechanisms already and are achieving latencies in the region of 2.5 - 3 microseconds. I say essentially, because I'm not convinced they've optimized quite to the degree I'm suggesting - reliable, scalable multicast RDMA isn't something you'll see a lot of even at a supercomputer fair. True, you're not getting that kind of latency over the Internet whatever you do, but if you can achieve a hard real-time guarantee of 3 microsecond delivery in a LAN party, you WILL notice a difference. At the very least, in the door price, which will now be expressed in exponential notation to fit on the door.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
If you see this and ponder buying this card for your game servers first try optimizing the Linux kernel. 1000 Hz ticks, big kernel lock preemption and other latency patches actually does wonders to ping times (and I do not mean ICMP echo or similar but ping packets answered by the game) and latency.
Many games have their own interesting capabilities for performance tuning. For instance Counter Strike 1.6 has the -pingboost setting which will switch between select() and alert() syscalls (10 ms reduction) or processing a frame for every packet. Other games have similar tuning options that will enhance performance. Then there's also tuning your network settings.
By the way, as far as I remember this Killer NIC is just some kind of offload engine. How *exactly* does this increase performance when most game specific packets are simple UDP packets that performance-wise are not as demanding as TCP packets (less checksums, no window scaling and other options easily tunable etc.)?
:/- spoon(_).
Aha, found a better source with some real info on this thing. Here.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Friends don't let friends enable ecmascript.
Actually, I found some more info on this after I made that post, and it's pretty interesting. Apparently there's more to it. First, they're bypassing the windows networking stack entirely, and that gives a measurable latency decrease. Second, the card is actually a linux machine on a card, so it runs a firewall and the windows drivers disable the windows firewall, and that gives another significant speed boost.
;)
Of course, all this presumes you're running windows to begin with, so it's useless to me. But for those so afflicted, perhaps a small subset are hyper sensitive to lag enough to make the thing be worth it. And it sounds interesting to me for other reasons - couldn't you stick a hard drive in directly attached to it and run bittorrent on your NIC?
=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Friends don't let friends enable ecmascript.
I believe the quote you are looking for is:
"Why does he have two clocks?"
"The one of the left is a spedometer for his extreme life. The other one is set to Tokyo time, so he knows when to drift race."
The $500 shoes worn by the professional will not be the same as the $500 shoes purchased by the average person. For one thing, the professional is paying for the technology and customization. The average person is paying for the marketing and endorsements.
That being said, the professional would NOT compare two shoes provided by a shoe company and "tested" on their own track.
S/He would compare them to his/her CURRENT favourite shoes on his/her current training track.
And that is where every single one of these KillerNIC "reviews" fails. It is not that difficult to swap a NIC. Yet the "testing machines" are always different. And none of the "reviewers" seem to be able to script a game. Or setup a test network with a test game server.
The "professional" in this case would setup a test network, with a test game server and a sniffer to see what is happening "on the wire" and script the game on his/her favourite machine with his/her current NIC.
Then the "professional" would swap the NIC's and re-set everything and run the script to see what difference/improvements there were.
It's not that difficult and it's not that expensive and yet not a single "review" of this "KillerNIC" seems to be able to do that.
Sure, you can pay $500 for shoes that were hand stitched by virgins under the light of a full moon with thread blessed by the Pope. And they may perform better than this other pair of shoes I'll give you to run in.
But in the end, you'd still be paying for the marketing of un-tested technology.
There's not much to offload in UDP. The checksums are about it.
*whoosh*
anand also did a review a month or so ago... what it amount to is that this is a consumer-marketed board that has server features enabled. Thats the same reason a gigabit Intel nic costs $400 when you can get "nearly" the same thing onboard for free. What that review amounted to was that it did work..but only in some cases. Much like expensive server Nics, if you got the other hardware for it it can help...but it's not "magic".