From cam.sri.com!larry Thu May 23 10:32:48 1996
Date: Thu, 23 May 1996 10:31:21 +0100
From: larry@cam.sri.com (Larry Piano)
To: johnca@cogs.susx.ac.uk
Subject: Re: tagger problems on sparc architecture
Cc: larry@cam.sri.com
X-Sun-Charset: US-ASCII

John,

It took me about 4 hours to track down the problems, but I think I've got a 
working version of label for snipe in ~ljp1003/mysntagger.  It won't run to 
completion on snipe because it runs out of file system space right near the 
end which makes no sense to me.  I believe it's writing the sec.ftr (J option, 
I believe) so maybe that can be dispensed with.  I ran out of time last night 
to test all the permutations.  It may be that there's something still wrong.  
Also, since snipe has only 36M of memory according to top, I had to run the 
tagger with only H3, k1 and K1.  The script I used to test snipe is in 
~/musnunk/myeval3.

I also copied the changes I made over to ~/mytagger and recompiled adder's 
version and it runs fine.  I copied label to ~/myunk as well.

If you can take the source in mysntagger, compile it at Sussex and get it to 
run to completion, let me know.  If not, I'll take another look at it.  If you 
want me to look at it again, could you have Graham give me login access to 
adder, nene or ouse from SRI?  It'd be a lot more convenient to do it from 
here ... .

Thanks,
Larry


From cogs.susx.ac.uk!johnca Thu May 23 16:46:39 1996
Date: Thu, 23 May 96 16:46 BST
From: johnca@cogs.susx.ac.uk (John Carroll)
To: larry@cam.sri.com
CC: johnca@cogs.susx.ac.uk
In-reply-to: <9605230931.AA09055@caernarfon.cam.sri.com> (larry@cam.sri.com)
Subject: Re: tagger problems on sparc architecture

Thanks for having a look at it. It seems to be better, though it's
still not at all happy running on sparcs. Here are the results of some
runs (rsune is the sparc I use here at Sussex). I made versions called
label.snipe and label.rsune on snipe and rsune respectively, and ftped
over label.snipe to run it here

rsune% label.snipe sustag B1 C2 H3 K3 k3 N tsec.trn dsec.lex msec.map O512
Segmentation fault

rsune% label.snipe sustag B1 C2 H3 K3 k3 J/tmp/jac N tsec.trn dsec.lex msec.map O512 | tail
,_,
loyal_JJ
husband_NN1
-_-
smack_NN1
!_!
-_-
and_CC
you_PPY
falr
Segmentation fault

rsune% label.rsune sustag B1 C2 H3 K3 k3 J/tmp/jac N tsec.trn dsec.lex msec.map O512
Bus error

rsune% label.rsune sustag B1 C2 H3 K3 k3 N tsec.trn dsec.lex msec.map O512
Bus error


snipe$ label.snipe ~/corpora/sustag B1 C2 H3 K3 k3 N J/anfs/tiptmp/jac/xx r/homes/jac/corpora/sec m/homes/jac/corpora/sec.map O512 | tail
good_JJ
,_,
loyal_JJ
husband_NN1
-_-
smack_NN1
!_!
-_-
and_CC
you_PPY
Memory fault

snipe$ label.snipe ~/corpora/sustag B1 C2 H3 K3 k3 N r/homes/jac/corpora/sec m/homes/jac/corpora/sec.map O512 
Memory fault


So only label.snipe runs, and only then if you use the J option, and
then it seems to crash almost at the end and the last couple of KB
don't get written out. label.rsune with the J option doesn't get as
far as writing anything to the J file

Could you have another look, please? I'll ask GT to add you to adder
and snipe from SRI

Thanks, John



From cogs.susx.ac.uk!johnca Tue May 28 13:53:47 1996
Date: Tue, 28 May 96 13:53 BST
From: johnca@cogs.susx.ac.uk (John Carroll)
To: larry@cam.sri.com
CC: johnca@cogs.susx.ac.uk
In-reply-to: <9605281059.AA24120@caernarfon.cam.sri.com> (larry@cam.sri.com)
Subject: Re: tagger problems on sparc architecture

>I believe I've got it but I can't get it to run all the way on snipe, because 
>there's not enough memory:
>
>
>    last pid: 23074;  load averages:  0.09,  0.16,  0.00            11:55:05
>    57 processes:  56 sleeping, 1 running
>    Cpu states:  0.6% user,  0.0% nice,  4.1% system, 95.3% idle
>    Memory: 36M available, 22M in use, 14M free, 2004K locked
>
>Not sure why there's 22M in use, because nether top nor ps -e will show me any 
>other user but I can't run.  Anyway, if you compile the source in ~mysntagger 
>on rsune, it should take you a bit further.  Let me know ...

Thanks for persevering. With the updates and a label binary made on
snipe I get similar symptoms both on snipe and rsune - Memory fault
before tagging gets underway, but I think after some lexicon
processing since it's not immediate. (So it doesn't get as far as last
time, when it bombed out just before finishing). I used the command
line

~ljp1003/mysntagger/label /homes/jac/corpora/sustag B1 C2 H3 K3 k3 N r/homes/jac/corpora/sec m/homes/jac/corpora/sec.map O512

(I think there's a different problem when I make binaries on rsune,
since even simple ones such as exdict fail straight away, whereas a
version of exdict built on snipe works fine. I'll get someone here to
look at it)

Can you have another go at label on snipe?

Thanks again, John


From cogs.susx.ac.uk!johnca Tue Sep  2 16:22:33 1997
Date: Tue, 2 Sep 97 16:23 BST
From: John Carroll <johnca@cogs.susx.ac.uk>
To: larry@cam.sri.com
Cc: johnca@cogs.susx.ac.uk
In-reply-to: <9605230931.AA09055@caernarfon.cam.sri.com> (larry@cam.sri.com)
Subject: Re: tagger problems on sparc architecture

Hi! Hope things are well with you and the family

Over the past couple of days I've finally got round to looking at the
problems the tagger has running here and I think I've fixed them. I took
the version you made for sparc just after you started at SRI. In case
you're interested I'll just briefly say what I did

1. the tagger as built here would hardly start up whereas one built on
snipe in Cambridge would get much further: fixed by stopping using the
special malloc library

2. an uninitialised variable in mainl.c was used in rare circumstances
before being set - the garbage value caused a segmentation error

3. for just 1 word in LOB, sprintf was being called with an argument NULL
for %s - the sparc didn't seem to like this but it was apparently OK on HP

I quite enjoyed tracking the problems down - never used gdb before. I also
had to compile it with with cc on HP so I could do a side-by-side
comparison of behaviour under gdb

Best wishes,

John




