[Vm-dev] Re: [squeak-dev] Add primitiveMillisecondClockMask? (was: The Trunk: Kernel-ar.564.mcz)

[Eliot Miranda]

On Sat, Apr 16, 2011 at 9:29 AM, Bert Freudenberg <bert at freudenbergs.de>wrote:

>
>
> On 14.04.2011, at 20:56, Eliot Miranda wrote:
>
>
>
> On Thu, Apr 14, 2011 at 3:23 AM, Bert Freudenberg <bert at freudenbergs.de>wrote:
>
>>
>>
>> On 14.04.2011, at 04:46, David T. Lewis wrote:
>>
>> >
>> > On Wed, Apr 13, 2011 at 07:07:57PM -0700, Colin Putney wrote:
>> >> On Wed, Apr 13, 2011 at 6:58 PM, David T. Lewis <lewis at mail.msen.com>
>> wrote:
>> >>
>> >>> Understood, though my question was intended to be about the existing
>> >>> primitiveMillisecondClock, which will presumably be retained for a
>> >>> few more years. Andreas committed a fix for a rather nasty bug related
>> >>> to a constant in the image being out of sync with a constant in the
>> VM.
>> >>> The method comment noted that "This should be primitive but it ain't"
>> >>> so I was just offering to add the primitive and update Squeak trunk
>> >>> to use it when available.
>> >>
>> >> Why does the existing primitiveMillisecondClock have to be maintained?
>> >> Is there a lot of code that depends on it rolling over? No sarcasm
>> >> intended, it's an honest question.
>> >
>> > It is used by Time class>>millisecondClockValue, so see senders of
>> > #millisecondClockValue.
>> >
>> >> From the point of view of the VM, the primitive needs to be maintained
>> > because many existing images expect it to be there. From the point of
>> > view of the image, a different and better implementation can easily
>> > be adopted. My question was from the point of view of the VM, if we
>> > assume that many images will expect primitiveMillisecondClock to be
>> > there, should we also add primitiveMillisecondClockMask such that
>> > those images can obtain the correct mask value if they choose to
>> > do so (both the VM and the image changes are minor updates).
>> >
>> > With respect to code that depends on the clock rollover, I expect
>> > that the most serious concern would related to socket timeouts, which
>> > might intermittently affect server applications, Seaside, etc.
>> >
>> > To be clear, the patch that Andreas committed will fully address
>> > the problem for all current VMs and images, so this discussion
>> > has little or no practical impact. It's really just about tidying
>> > up some loose ends just in case the the primitiveMillisecondClock
>> > ends up getting used for a few years longer than any of us might
>> > care to anticipate ;)
>> >
>> > Dave
>>
>> I agree that it should have been made a primitive long ago. However,
>> adding it now would only make sense if we anticipate that it will ever
>> change. And since Eliot is advocating to switch over to a 64 bit clock
>> anyway, what's the point of changing the old one now?
>>
>> I'd still retain the old primitive for accessing millisecondClockValue
>> because it certainly can be cheaper (no LargeInts). Or was there a proposal
>> to actually change the old primitive, instead of adding the high-res one?
>>
>
> I've heard this argument advanced (that the old interface is cheaper
> because of no large integers) but I think it's largely a straw man.  If you
> consider the simplicities that accrue from using a 64-bit clock that can't
> wrap and is synchronised to wall time (i.e. that Time now is derived from
> those same microseconds) then it's reasonable to assume the large int
> allocation cost will be substantially if not entirely recouped by
> simplicities higher up in Time, Delay et al.  No, I don't have figures; only
> experience with VW where we didn't notice any degradation but did loose a
> number of painful 45 1/2 day bugs.
>
> 2¢
> Eliot
>
>
> I don't disagree, in particular with running on Cog which is much more
> comparable to VW than the interpreter. But what speaks against retaining the
> old primitive even if it is unused in modern images?
>

Nothing.  I just ant to get us onto a good time basis, and I need to stop
blathering and do something :)



> So do you propose to add the primitiveMillisecondClockMask primitive in
> anticipation of changing primitiveMillisecondClock to answer an absolute
> time?
>

I wouldn't do that.  In Cog (and these can be added to Interpreter) there
are separate primitives that answer UTC and local microseconds since 1901.
 The old wrapping milliseconds is still available from the old primitive but
implemented by truncating the 64-bit microseconds (IIRC after subtracting
the start time)


>
> In light of what Igor reported I'd think that having the time be
> monotonically increasing is quite valuable, too. Or do we need to have
> in-image logic to catch the clock jumping backwards?
>

We had a discussion a few months ago about the right way to handle this in
the VM, i.e that the VM assumes that jumps of the order of a few minutes are
clock adjustments, not drift. I need to implement this and test against
adjusting the clock backwards.  How Delay copes with clock adjustments is
another issue, both for durations and absolute deadlines.  What's a
reasonable trade-off between implementation feasibility, desired semantics
and sanity?  Right now Delay provides only durations (seems fine to me) but
under the covers this gets turned into a deadline.  What happens to a
duration (and under the covers to a deadline) when the clock is adjusted?


> - Bert -
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20110416/ae4f32cb/attachment.htm