Re: What's the slowest SPARC box that can saturate a gigabit link?

From: Jorgen Moquist (jorgen.moquist_at_n.o.s.p.a.m.mailbox.swipnet.se)
Date: 12/25/04

  • Next message: te_at_ivorypetal.com: "Sun PCI QFE Card..." 
    Date: Sat, 25 Dec 2004 00:40:29 +0100

    dion_b wrote:
    > CJT wrote:
    >
    >> Jorgen Moquist wrote:
    >>
    >>> CJT wrote:
    >>>
    >>>> What's the slowest SPARC box that can saturate a gigabit link
    >>>> (assuming suitable disks to provide data at the requisite rate)?
    >>>>
    >>>> Will a lowly IIi at, say, 333 MHz (as might be found in an Ultra 5)
    >>>> do it? Or will it run out of horsepower?
    >>>>
    >>> read somewhere at docs.sun.com that 100mbit network requires 110MHz cpu.
    >>> so 1000mbit would use 1100Mhz cpu. ( and i think backplane to ! )
    >>> can the pci on a on U5 really transport that amount of data ?
    >>> /Jörgen
    >>>
    >>
    >> It may well be that the pci bus in the U5/U10 is the limiting factor;
    >> that kind of insight is what I'm seeking.
    >>
    >> If the alleged (haven't seen them myself yet) improvements of S10 and,
    >> say, a 450 MHz CPU would get me close, that would be significant.
    >
    >
    > Dead right regarding the PCI bus. The U5/U10 have a 33MHz, 32b bus that
    > can do a theoretical max of 133MB/s, which is just about a Gb. However
    > you must remember that every other device in the system uses that bus,
    > in particular the hard drive controllers that would supply the NIC with
    > data.
    > Plus no PCI bus controller chipset ever reaches the theoretical maximum.
    > I have never seen a benchmark of the U5/U10, but on x86 PCs 110MB/s is
    > the absolute maximum and 90MB/s is closer to the norm.
    >
    > Unless you are operating purely from RAM (which does not use the PCI bus
    > of course), reckon on an equal amount of bandwidth used for drives and
    > for NIC, which means 45MB/s is about the theoretical maximum, a 'mere'
    > 360Mb/s
    >
    > If you want to saturate that Gb, you need 64b PCI, preferably at 66MHz
    > or above, or you need the NIC and drive controllers on seperate PCI buses.
    >
    > Even if you're working from RAM you will never totally saturate a Gb
    > link with 32b 33MHz PCI, although you may come close enough for your
    > purposes if they are for a purely synthetic test

    I once did a trial installation for a U10/quad qfe card sunray server.
       3 terminals 3 remote stations.
       using 2 ports on the qfe plus the hme0, the customer was dissapointed.

       then changed to an E450 (about same mhz)qfe was now on 64bit slot.
       and system disks on different bus (scsi) and success !
       customer now has bought +90 workstations 6 servers lots terminals.

       the U5/U10 pci-bus cant
       drive a quad qfe card fully on 32bit and worse while using ide disk.

       a scsi-hba on a U5/U10 will make huge difference to performance
    /Jörgen

  • Next message: te_at_ivorypetal.com: "Sun PCI QFE Card..."