Re: Pathological behaviour when building large hash index

6 Dec 2013

      w.r.t. the slow hash index creation, I did find an OS setting that improved
things dramatically - first hash building query is now ~3 minutes rather
than 4.5 hrs.

I found that my /proc/sys/vm/dirty_background_ratio setting was 0, and so
the OS kicked in and started writing as soon as pages were dirty.  This
gets in the way of the page fault servicing.  I set it to 20, which now
means that the OS doesn't start flushing for a while and so mserver gets to
max out CPU dirtying up pages during the thash creation.  A little bit
later, the flusher kicks in and mserver slows down to 0.1 CPU.  Flushing
completes, mserver picks up to 100%, etc...  We make a lot quicker progress.

I'm still confused by why my OS doesn't seem to want to keep all the pages
of thash in memory.  If I look in /proc/{pid}/numa_maps I can see that
during the thash creation, for the thash mem mapped file the 'active' page
count (which I believe is the memory resident portion) seems to not want to
increase to more than ~0.25 of the total 'mapped' page count.  My
assumption is that if this proportion was closer to 1.0, then my memory
access wouldn't be causing page faults in the first place.  I've got lots
of RAM spare, and am using interleaved numa settings, so I don't understand
what's preventing active growing.  Does anyone have an idea, or know how I
can diagnose this?

-Will

On 5 December 2013 19:57, Will Muldrew  wrote:
...
No index!
On 5 Dec 2013 17:05, "Stefan Manegold"  wrote:
...
hm, that's with the index in place?
If so:
good for our optimizer,
but unfortunately not so good for you ...
If not:
- good for you!
- impressive for our "execution engine" (scan)
(- "embarrassing" for our optimizer ... well, it's a minor detail ...)
;-)
Stefan
...
Sorry!
sql>select count(*) from t_order2 where orderId between 'AAAAAA' and
'AAAAAA';
+------+
| L1 |
+======+
| 0 |
+------+
1 tuple (596.202ms)
On 5 December 2013 16:49, Stefan Manegold < Stefan.Manegold@cwi.nl >
wrote:
Hi Will,
thank for reminding me: the optimizer still sees your query as point
query
and thus prevents parallelization in order to build/use the hash index.
Maybe you can "fool" the optimizer (and I hope noone else is reading
what I'm
writing now ... ;-)) by "disguising" the point predicate as range
...
say, "where orderId between 'FOO' and 'FOO'" --- in fact, I rather hope
(for
us)
that is will not work ...
Stefan
----- Original Message -----
...
Hi Stefan
I've added that hash = FALSE patch, and now my simple query returns
reliably
in 6 seconds. It's still a lot slower than the indexed one
(obviously), and
12x slower than the one with the time > '1970-01-01' optimiser nudge
(which
causes parallelism), but it'll do for now.
Thanks
-Will
On 5 December 2013 16:32, Will Muldrew < will.muldrew@gmail.com >
wrote:
Hi Stefan
Here are my version details:
MonetDB-11.15.17
CentOS release 6.3
kernel 3.6.2
I think I might proceed with a patched version with the hash disabled.
That's
totally fine for me right now, though obviously it'd be nice to be on
...
...
trunk!
W.r.t. a less random access hash generation. This is all very
hand-wavey
but
how about: 1) generate a sequence of (hash,oid) pairs to some file or
mmapped area. 2). use some well-known "external sort" algo to sort
...
...
pairs by hash (e.g. split, qsort, merge). 3). build your final hash by
iterating through this sorted datastructure.
Your memory writing would then be sequential and you'd get good page
hit
characteristics. You just need to choose a decent sort algo - ideally
which
works transparently over memmapped data.
What do you think? I might even have a crack at it myself :)
-Will
On 5 December 2013 15:54, Stefan Manegold < Stefan.Manegold@cwi.nl >
wrote:
Hi Will,
could you please also share with us which version of MonetDB you are
using,
and OS you are running?
Pleass find more comments & questions inlined below.
----- Original Message -----
...
Hi
I've been investigating very slow queries on my 350M row dataset
when
selecting for equality on a non-unique string id. e.g.
select count(*) from t_order where orderId = 'FOO';
Is your orderId column unique, or are there duplicate values?
...
This query takes ~4.5 hours to build a thash index. Subsequent
repeats
hit
the cache and are sub milli.
select count(*) from t_order where time > '1970-01-01' and orderId =
'FOO'
This second one is consistently finished in a few seconds or less,
and
doesn't create any indexes.
The first query is a bit of a show-stopper. It spends a huge about
of
time
in
BAThash (gdk_search.c), and incurs a lot of page faults while
reading and
writing at random across a multi GB mmaped thash file. Additionally
...
...
...
page faults are hampered by the OS trying to write dirty pages out
in the
background. I've got plenty of RAM overall (130GB), but the
'active' page
proportion for the thash file seems to be stuck at about 25% -
giving me
a
0.75 probability of a miss.
* Is there some way to make my OS less keen to evict pages?
(swappiness =
0
already) I should have plenty of room to have the whole thash file
resident.
Additionally, allowing it to sit on dirty pages longer would reduce
...
...
...
total write IO.
In fact, if there is sufficient memory in total, there would be no
need to
flush
dirty (non-persistent) pages at all, but we have not yet found a good
way
to
hit
this to the OS --- madvice() did not have much/any useful efficet for
us on
Linux.
Maybe some of the issues dicsussed here play a role:
http://engineering.linkedin.com/performance/optimizing-linux-memory-manageme...
...
?
While we have not yet encountered the NUMA problem reported there,
we have encountered the transparent huge page problems, and usually
disable them as also suggested in the above posting.
...
* Is there some way to simply prevent large thash creation? For many
applications I'd rather have slower consistent queries, then incur a
massively slow query following a restart (not great in a prod
environment!)
Not easily right now.
However, you can either add the "fake" time predicate as you did
above,
which (in this case) triggers MonetDB's optimizer to push the time
predicate
(either because it's in timestamp (int) rather then varchar (string),
or
because the time column happens to be ordered) below the orderId
...
...
and this avoid the hash build,
or change the following line(s) in gdk/gdk_select.c (in case you built
MonetDB yourself from source):
hash = b->batPersistence == PERSISTENT &&
(size_t) ATOMsize(b->ttype) > sizeof(BUN) / 4 &&
BATcount(b) * (ATOMsize(b->ttype) + 2 * sizeof(BUN)) <
GDK_mem_maxsize / 2;
->
hash = FALSE;
We'll consider whether we can provide a better solution in a future
release.
...
* Is there some way to generate the hash index with a more
sympathetic
algo
that doesn't degrade so steeply. e.g some hash + sort. (I have
----- Original Message -----
predicate,
the
these
these
the
predicate
literally
...
...
...
no
experience with this!)
We'll also look into this. However, building a hash index inherently
incurs
random access, so chances are slim ...
Best,
Stefan
...
-Will Muldrew
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list
--
| Stefan.Manegold@CWI.nl | DB Architectures (DA) |
| www.CWI.nl/~manegold/ | Science Park 123 (L321) |
| +31 (0)20 592-4212 | 1098 XG Amsterdam (NL) |
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list
--
| Stefan.Manegold@CWI.nl | DB Architectures (DA) |
| www.CWI.nl/~manegold/ | Science Park 123 (L321) |
| +31 (0)20 592-4212 | 1098 XG Amsterdam (NL) |
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list
--
| Stefan.Manegold@CWI.nl | DB Architectures   (DA) |
| www.CWI.nl/~manegold/  | Science Park 123 (L321) |
| +31 (0)20 592-4212     | 1098 XG Amsterdam  (NL) |
_______________________________________________
users-list mailing list
users-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/users-list