The dream platforms for Dragonfly?

Matthew Dillon dillon at apollo.backplane.com
Sat Sep 25 13:06:08 PDT 2004


:How are you thinking you solve the latency issues (and QoS) on a WAN e.g.?
:A lot of experiments and research will be burned on a scheduler that
:can do these things on a fast reliable LAN alone.

    The only way to deal with latency issues is to implement a fully
    coherent cache subsystem, e.g. MESI or something like that.  With
    a general infrastructure in place it becomes an issue of what kind of
    algorithm is being used and the clustering topology constructed.
    Scheduling can be a complex issue but most seriously distributed 
    problems are hand-tuned anyway, or placed in cookie-cutter groups, so
    scheduling is actually less of a problem for those.

    A general clustering architecture is not meant to 'solve' the latency
    issue per say, but it IS meant to allow the sysop to construct whatever
    kind of topology he or she wants (LAN, WAN, hybrid, normal network links,
    multi-master SCSI, other types of transport)... i.e. to make use of
    whatever is available.

:Dunno but i get a wierd feeling when ppl start asking what HW they shall
:buy for something i imagine like 5 years away and with only a certain
:probability of success. So i think the _long time_ part of the project
:should be more stressed and for the short term the usual hardware
:should do.
:
:Andy

    I never buy leading-edge hardware, simply because I would be paying a 
    premium for performance that will become old news a year from its purchase.
    It just isn't worth it.

    Right now I have two AMD64 based cubes (Shuttle SN85G4 case, Athlon 3200+
    cpu, etc), one AMD64 based tower (Asus K8V MB, Athlon 3200+ cpu, which
    is now LEAF), and a couple of Shuttle SN41G2 based AMD Athlon XP 2600+
    based cpus.

    All the DragonFly production machines, except LEAF, are now running
    on SN41G2/XP-2600+'s with lots of disk and 1G of DDR 333 memory. 
    They seem give the best bang for the buck, far faster (and run cooler) 
    then P4 celeron boxes yet not overpowered for their designated tasks.
    A typical configuration is:

	* Kingston PC2700 DDR333 512MB		$90	x 2	(up to 2)
	* Intel PWLA8391MTBLK PRO/1000 (em0)	$40	x 1	(optional)
	* Shuttle SN41G2 Version 2 barebone	$270	x 1
	* Sony 52x internal cd-rom		$16	x 1
	* AMD Athlon XP 2600+
	  (1.917GHz, 512KB, 333MHz, Socket A)	$94	x 1
	* Maxtor DiamondMax plus9 160GB 	$105	x 1	(up to 2)
	  ATA133 7200 rpm 8MB IDE HD

	    Total:	$705 (1G of memory)
		 or	$575 (512M of memory, use built-in 100BaseT
				ethernet only instead of GiGE)

    I really like the SN41G2 Cube enclosures.  They have built-in everything
    (video, ethernet, usb, firewire).  They have one PCI slot and one AGP
    slot.  I usually throw an Intel PRO/1000 GigE card in the PCI slot.  Even
    more importantly, they did a real good job on the internal layout.  The
    CD/HD bay is top-loaded and easy to pull out, and the cable routing is
    reasonable.  There is room for a CD and two HDs (though I will note that
    packing that much junk into a cube requires that you set the Fan to max).
    The only thing they do not have is SATA.

    I also have two 2xP3 cpu smp DELL-2550's left over from earlier years 
    and a few other assorted boxes.  I used to have a lot more DELLs but I've
    gotten rid of most of them.  The two DELL-2550's I have left are 
    relegated to being SMP test boxes.  I am not going to purchase any new 
    SMP boxes until the multi-core Opterons come out, I just don't see the
    point.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>





More information about the Kernel mailing list