In message <Pine.NEB.4.61.0502151039020.22835@chylonia.3miasto.net>, Wojciech P
uchar writes:
>> left.
>>
>> Sure, you can eliminate the CPU usage with iSCSI by purchasing an expensive 
>> iSCSI offload adapter, but now where is the cost savings?  They go "*poof*",
> 
>> as you observe.  Sure, you save on the FC switch, but a high-end Gig-E switc
>h 
>> that can support jumbo frames and traffic shaping ain't exactly chopped
>
>could you please explain why to "save costs" instead of just not making 
>costs = not introducing extra architecture, extra protocol, extra new 
>network standard?
>
>NFS gets 10MB/s with pentium 100 machine with less than 50% cpu loaded, at 
>least with NetBSD, without "intelligent" network adapter.
>
>it will be much less when (as UVM promise) memory-to-memory copy with NFS 
>traffic will be eliminated, and i'm sure someone will do it.
>
>i'm talking about such low end (for todays standards) machines like 
>pentium 100, with P4 CPU load with 100MB/s NFS shouldn't be noticable.
>
>and there is really rare case there are 100MB/s disk traffic...
>
>
>anyway would be nice to this ethernet-connected drives to have NFS support 
>too not only windows CIFS.
>
The original proponents of this technology were looking at high-end 
systems -- they'd laugh at the thought of merely 10 MB/s.  When they 
came to the IETF for standardization and I was brought in to talk about 
security, the issue was whether or not IPsec could run at GigE or 
10GigE speeds.  They couldn't use 3DES, not because it's insecure, but 
because they'd have to rekey every 2.5 minutes at GigE rates and every 
15 seconds for 10GigE.  Fortunately, AES's 128-bit blocksize solved 
that problem.

On the other hand, they weren't even convinced they needed crypto at 
all -- they had a data center model in mind, and didn't foresee (or 
believe in) building area connectivity, let alone WAN connectivity, to 
such boxes.

Why were they pursuing this?  In a word, cost.  I can buy GigE boards, 
quantity 1, for US$25, and a GigE switch for $50.  I doubt I have to go 
up by more than a factor of 10 to actually get that throughput from the 
cards...  10GigE is still expensive, but it won't be for long.  In 
other words, the hardware is *much* cheaper than FC, and it's riding a 
much steeper price/performance curve.  You also get away from needing 
custom device drivers for the disk attachment, since the host vendor 
and/or the Ethernet board vendor will supply those.  

Of course, you also end up stressing other parts of the total system; 
we've already heard about RDMA (though I'm frankly very skeptical; I 
don't think they understand the real bottlenecks yet, nor the full 
implications of RDMA).

In short -- don't assume that your needs are everyone else's.  You -- 
and for that matter I -- may not need really high bandwidth with a disk 
attach model.  Some people do -- and the question is whether or not the 
rest of us can benefit from their technology, once the cost comes down 
enough.  

		--Prof. Steven M. Bellovin, http://www.cs.columbia.edu/~smb