Very strange measurements

[Brecht Vermeulen]

Benjie,

thanks for looking into it.

The tulip card was a DLINK 4-port DFE-570 card, but I used *exactly* the
same card in the K6-500 MHz machine as in the Athlon 1 GHz machine, so
this can't be the problem I think.

Could it help if I gave you remote access to the machine ?

Can there be a problem with motherboards or something like that for this
problem ?

regards,
Brecht

Benjie Chen wrote:
> 
> i am not sure what's going on. it seems like the device is claiming
> to be busy in the slow case more often the fast case, which is weird.
> 
> what kind of tulip card were these again? may be we can purchase one
> here and try it out.
> 
> benjie
> 
> On Sat, Jul 07, 2001 at 10:39:08AM -0700, Eddie Kohler wrote:
> > Here is a summary of Brecht's data.
> >
> > POLLDEVICE                 "Slow" case     "Fast" case     Ratio
> > ========================     -----------     -----------     -----
> > Poll cycles/scheduling           101.305        60.157      1.68
> > Packets/activation                 1.000         1.002      1.00
> > Refill cycles/activation          96.418        73.514      1.31
> > Alloc cycles/activation          313.577       136.413      2.30
> > Push cycles/activation         45048.118      3688.913     12.21
> > Total schedulings            149583701        77971748      1.92
> >
> > TODEVICE                   "Slow" case     "Fast" case     Ratio
> > ========================     -----------     -----------     -----
> > Dev busy/activation              114.664           0.057  2012.239
> > Hard start/activation              0.096           0.000  3192.317
> > Pull cycles/activation           181.176          98.689     1.836
> > Clean cycles/activation          664.575         334.151     1.989
> > Free cycles/activation           617.954          77.850     7.938
> > Queue cycles/activation          267.944         131.332     2.040
> > Activations                        83639         1034375     0.081
> >
> > Note: Benjie, we should think about redesigning the stats. There is no way
> > to determine how many times ToDevice was scheduled, for example.
> >
> > Most intriguing to me are PollDevice push cycles/activation, and ToDevice
> > dev busy/activation and hard start xmit/activation. How can there be 100+
> > dev busys per activation!?!??!?!?!? Shouldn't there be at most one? (An
> > "activation" is a period where the ToDevice sent a packet: `if (sent > 0)
> > _activations++;')
> >
> > love,
> > ed
> 
> --
> benjie chen
> benjie at lcs.mit.edu