[Beowulf] DIMMnet-1 DIMM slot NIC?
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Joachim Worringen joachim at ccrl-nece.deTue Jul 27 01:11:08 PDT 2004
- Previous message: [Beowulf] Intel microcode updates for "erratum" - "an incorrect instruction stream may be executed"
- Next message: [Beowulf] IBM gets nod to build supercomputer
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Bill Broadley wrote: > On Mon, Jul 26, 2004 at 06:26:59PM +0200, Joachim Worringen wrote: > >>Keith D. Underwood wrote: >> >>>Now, if you want to implement something like UPC or SHMEM and you have >>>all of the shared address space located on the card plugged into that >>>slot (maybe not incredibly practical, but probably possible) then a DIMM >>>interface may be your best bet. >> >>Even then, you have the problem that this approach doesn't scale: a >>single memory slot can cover only a small fraction of the memory of the >>N other nodes (think of N >> 10), which requires to fall back on message > > What is your reasoning? Bandwidth limitations? Address bits? Latency? As I wrote, not enough address space. A single DIMM can cover at most 50% of the address space of the local node, and never more than the local node may have physiscally installed. But the accumulated memory of all nodes in a cluster is much larger, and thus the fraction of remote memory that a node could access directly through such a NIC would decrease as the number of nodes in the cluster increases. Thus, if you have 100 nodes, you can only access 1% of each nodes memory directly. This may be sufficient for message passing, but not for shared-memory programming models, or even RDMA. Joachim -- Joachim Worringen - NEC C&C research lab St.Augustin fon +49-2241-9252.20 - fax .99 - http://www.ccrl-nece.de
- Previous message: [Beowulf] Intel microcode updates for "erratum" - "an incorrect instruction stream may be executed"
- Next message: [Beowulf] IBM gets nod to build supercomputer
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
