🚀 go-pugleaf

In article <321@tlvx.UUCP> sysop@tlvx.UUCP (SysOp) writes:
>kent> [...] the size of gates in modern memory is much smaller and
>kent> thus they are more susceptable to alpha radiation induced parity
>kent> errors than were the gates of the Cray memory.
>
>Even "more susceptable"?  Ok, why?  How?  If it were really that bad, then
>people using the A3000 now should be at least occasionally noticing weird
>things happening, right?  But are they?  What about my A1000 with 2.5 megs?
>(Of course, I don't have parity, so how can I tell?  Sigh.)

Sorry; its amazing how susceptible I am to thinking that just because I
have known something for a decade or more, that it therefore is common
knowledge.  Herewith the (stupidly) omitted explanation:

Alpha radiation (a fast moving, stripped helium nucleus) originates within
the naturally occuring radioactive impurities of the memory chip itself.  By
their nature (big, bumbling and slow compared to other kinds of radioactivity)
alpha particles have very limited penetrating power; they do all their
mischief near their point of origin.  For our purposes, the thing of interest
about their action is that, being an extremely positively charged particle
among atoms essentially in neutral balance, they have a large effect on
the outer shell (conduction) electrons, pulling large numbers after them in
their wake.  They cause a parity error when these entrained electrons are
deposited in a spot that causes a gate to shift its state from 0 to 1 or
vice versa, for instance on one of the control lines.

The older memory chips had very little susceptability to alpha radiation
induced parity errors.  Although the alpha radiation exists constantly
at a low level in every chip, "large number of electrons" above must be
considered relative to the number of electrons required to switch a gate.
The older memory chips, with larger "wiring" and component sizes, used
larger switching currents; significantly larger than the amount of charge
moved by one alpha particle.

Since dynamic RAM means the memory is refreshed repeatedly by renewing the
control charges holding the state (0 or 1) of each gate, there is not usually
time for the charges carried by individual alpha particles to accumulate
from several events to switch a gate, before the refresh cycle sets the charge
back to its nominal value.

In contrast, in denser, newer memory chips with smaller "wiring" and
components, the charge delivered by a rogue alpha particle is of
comparable size to the holding charge on a gate, and so the gate may
be switched before a refresh cycle can correct the problem.  Making
the refresh cycles faster (than they are, not than the old circuits) is not
an option, because most computer chips these days are heat limited, and
more refreshes means more heat.

So for an individual gate, denser memory means a larger chance of a bit
being flipped _from_this_one_cause_.  Still, chips are not highly
radioactive, so for a single bit, this is a very low probability.

The problem comes when you accumulate megabytes of these bits together;
the chances of all of them avoiding errors tail off rapidly as their
number increases, in math similar to that  the birthday paradox employs.

I'm a bit shakey on the numbers here, since I was last a hardware
practicioner in 1972 and things have changed a trifle, but to the
best of my understanding, with today's component sizes, speeds, and
numbers of megabytes, you can expect to get in trouble somewhere
between 1 and 100 megabytes.  I defer to today's hardware practitioners
for better data.

As to why you don't see problems in your 3Meg AT, well, for one thing, as
you mentioned, you don't have parity checking, so they could get by.
Next, most of the software you run (or at least what I ran when using a
5 Meg '386 box) is unused by most applications, still stuck at the 640K
limit.  Again, at least in my Amiga, about 5 megabytes is loaded with
software I may not use from boot to boot, but keep around because it is
convenient.  More, in running code lots of the code space is never
touched (use a file zapper; lots of it is huge blocks of zeros).  Again,
stuff such as screen memory, if you get a bit flipped, you may never
notice before you switch screens or windows in a screen and rewrite the
soft parity error with good data.  Similarly, would you really be likely
to notice a one bit error in a sampled sound data block?

Besides the above, your machine may sit idle 20 hours a day, not even
powered up.  In summary, there are lots of reasons why alpha induced
parity errors would not be a big enough problem to become noticable.

Yet.

But like the birthday paradox, you don't have too far to go in terms of
bigger applications exercising more of the machine, full time unattended
operation (e.g. raytracing, doing accounts), more memory, more critical
applications, and so on, before you run into Seymore Cray's problem.
Parity checking is a necessity in large machines, just to be able to
rely on the results the machine gives you.  Error correcting circuitry
is a necessity in large machines, to get the kind of uptime and through-
put the machine's raw speed and memory size seem to promise.

That's probably more than you wanted to know, and please excuse any
details that might not be "just so".  Since I stopped doing this stuff
for a living, I'm a fairly casual student of the art.  More is available
in IEEE pubs, Scientific American, and so on.

Kent, the man from xanth.
(xanthian@zorch.sf-bay.org)