EXPLAIN ANALYZE printing logical and hardware I/O per-node

On Fri, 2007-12-14 at 15:47 +0000, Gregory Stark wrote:

I've wanted for a long time to have EXPLAIN ANALYZE output per-node I/O usage.
This would be especially useful if we could distinguish hardware versus
logical I/O though. And I always thought that would be very hard.

My thought in the past was that would could do it on Solaris by having
Postgres use DTrace directly via its (undocumented but existing) programmatic
interface.

For other operating systems it was tempting to suggest just timing the read(2)
call to see if it took too long to be a logical operation. The problem there
is that gettimeofday would impose far too much overhead to make that practical
(or even be precise enough to work properly).

But it occurred to me just now that the hardware instruction counter available
on just about every platform would be good enough for a heuristic guess at
whether the read(2) was cached.

I'm skeptical that this would be reliable enough to be very useful,
especially in the face of concurrent, unpredictable system activity on a
busy system. I agree that it would be useful information, though.
Perhaps a useful first step would be to teach EXPLAIN ANALYZE to report
the number of logical and physical I/Os from Postgres' perspective (i.e.
physical I/O just means we need to go to the kernel).

Hi,
I already made a discussion about it. We can view the Logical I/Os. If
we enable the log_statement_stats in the conf file and apply the following
patch, it is possible. But putting it in Explain analyze makes more sense to
me.

*** postgresql-8.3beta1/src/backend/storage/buffer/bufmgr.c     Tue Sep 25
18:11:48 2007
--- postgresql-8.3patch/src/backend/storage/buffer/bufmgr.c     Fri Oct 19
23:18:36 2007
***************
*** 1470,1477 ****
               localhitrate = (float) LocalBufferHitCount *100.0 /
ReadLocalBufferCount;

       appendStringInfo(&str,
!       "!\tShared blocks: %10ld read, %10ld written, buffer hit rate =
%.2f%%\n",
!                               ReadBufferCount - BufferHitCount,
BufferFlushCount, hitrate);
       appendStringInfo(&str,
       "!\tLocal  blocks: %10ld read, %10ld written, buffer hit rate =
%.2f%%\n",
                                        ReadLocalBufferCount -
LocalBufferHitCount,
LocalBufferFlushCount, localhitrate);
--- 1470,1477 ----
               localhitrate = (float) LocalBufferHitCount *100.0 /
ReadLocalBufferCount;

appendStringInfo(&str,
! "!\tShared blocks: %10ld Logical Reads, %10ld Physical Reads, %10ld
written, buffer hit rate = %.2f%%\n",
! ReadBufferCount, ReadBufferCount -
BufferHitCount,
BufferFlushCount, hitrate);
appendStringInfo(&str,
"!\tLocal blocks: %10ld read, %10ld written, buffer hit rate =
%.2f%%\n",
ReadLocalBufferCount -
LocalBufferHitCount,
LocalBufferFlushCount, localhitrate);

--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Group.
(www.alliedgroups.com)

"Gokulakannan Somasundaram" <gokul007@gmail.com> writes:

Hi,
I already made a discussion about it. We can view the Logical I/Os. If
we enable the log_statement_stats in the conf file and apply the following
patch, it is possible. But putting it in Explain analyze makes more sense to
me.

I was going to say that I'm really only interested in physical I/O. Logical
I/O which is satisfied by the kernel cache is only marginally interesting and
buffer fetches from Postgres's shared buffer is entirely uninteresting from
the point of view of trying to figure out what is slowing down a query.

However I suppose that's not true. There are other reasons why buffer fetches
could be interesting. In particular I imagine when users post explain analyzes
it would give us a good idea of whether their tables or bloated or their
tuples are extremely wide (in cases where the planner gets it wrong).

But I do think that showing logical I/Os without even an heuristic based
measurement of actual physical i/o is pretty useless. It will make people
think they want to grow their shared buffers to cover all of memory.

--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's PostGIS support!

I was going to say that I'm really only interested in physical I/O. Logical

I/O which is satisfied by the kernel cache is only marginally interesting
and
buffer fetches from Postgres's shared buffer is entirely uninteresting
from
the point of view of trying to figure out what is slowing down a query.

Ok the Physical I/Os are already visible, if you enable log_statement_stats.
Again i accept that it would be more helpful, if it gets displayed with
Explain Analyze.

However I suppose that's not true. There are other reasons why buffer
fetches
could be interesting. In particular I imagine when users post explain
analyzes
it would give us a good idea of whether their tables or bloated or their
tuples are extremely wide (in cases where the planner gets it wrong).

I have used it a lot for query tuning. If we re-write a query in such a way,
the logical reads will come down, then it implies lesser physical reads in
production. I think you would accept that there are some ways in which the
query can be re-written only by humans and not by the optimizer. When we do
that, instead of looking at the explain analyze time, it makes more sense
for me to look at the logical reads

But I do think that showing logical I/Os without even an heuristic based
measurement of actual physical i/o is pretty useless. It will make people
think they want to grow their shared buffers to cover all of memory.

I just want to clarify that we should display both Logical reads and
physical reads together. But increasing the shared buffer by looking at the
performance of a query doesn't seem to be a good idea. But people should be
aware that Logical reads is not for shared buffer management.

--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's PostGIS support!

--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Group.
(www.alliedgroups.com)

Gokulakannan Somasundaram wrote:

I was going to say that I'm really only interested in physical I/O. Logical

I/O which is satisfied by the kernel cache is only marginally interesting
and
buffer fetches from Postgres's shared buffer is entirely uninteresting
from
the point of view of trying to figure out what is slowing down a query.

Ok the Physical I/Os are already visible, if you enable log_statement_stats.

I think you missed the point. What log_statement_stats shows are not
physical I/Os, they're read() system calls. Unfortunately there's no
direct way to tell if a read() is satisfied from OS cache or not. Greg's
suggestion was about how to do that.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

On Dec 16, 2007 1:03 AM, Heikki Linnakangas <heikki@enterprisedb.com> wrote:

Gokulakannan Somasundaram wrote:

I was going to say that I'm really only interested in physical I/O.

Logical

I/O which is satisfied by the kernel cache is only marginally

interesting

and
buffer fetches from Postgres's shared buffer is entirely uninteresting
from
the point of view of trying to figure out what is slowing down a query.

Ok the Physical I/Os are already visible, if you enable

log_statement_stats.

I think you missed the point. What log_statement_stats shows are not
physical I/Os, they're read() system calls. Unfortunately there's no
direct way to tell if a read() is satisfied from OS cache or not. Greg's
suggestion was about how to do that.

Oh OK. Thanks for clarifying..

--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Group.
(www.alliedgroups.com)

On Dec 14, 2007, at 11:10 PM, Neil Conway wrote:

But it occurred to me just now that the hardware instruction
counter available
on just about every platform would be good enough for a heuristic
guess at
whether the read(2) was cached.

I'm skeptical that this would be reliable enough to be very useful,
especially in the face of concurrent, unpredictable system activity
on a
busy system. I agree that it would be useful information, though.
Perhaps a useful first step would be to teach EXPLAIN ANALYZE to
report
the number of logical and physical I/Os from Postgres' perspective
(i.e.
physical I/O just means we need to go to the kernel).

*watches hands wave*

If we assume that what we *really* want to see the difference of is a
cache IO vs one that truly hits a platter, I don't see how you could
come up with enough variation to account for that. Unless you're
super lucky and the sector you need is just about to hit the head
when you issue the request, you're going to have a delay measured in
milliseconds. 1ms on a modern CPU is over 100,000 cycles. That's a
lot of instructions...

Someone want to throw together some code that actually measures this?
Maybe something that keeps a histogram of how many instructions take
place per I/O request? If it turns out that counters do vary too much
between CPUs, there might be ways that we can account for that.

Also, has anyone looked into adding a class of system calls that
would actually tell us if the kernel issued physical IO? I find it
hard to believe that other RDBMSes wouldn't like to have that info...
--
Decibel!, aka Jim C. Nasby, Database Architect decibel@decibel.org
Give your computer some brain candy! www.distributed.net Team #1828

On Mon, 17 Dec 2007, Decibel! wrote:

Someone want to throw together some code that actually measures this? Maybe
something that keeps a histogram of how many instructions take place per I/O
request? If it turns out that counters do vary too much between CPUs, there
might be ways that we can account for that.

I'd expect the easiest way to do a proof of concept here would be to use
the Linux oprofile tool. That's already abstracted away the differences
in hardware counters and provides a relatively simple interface to collect
the data without getting bogged down with implementation there. Starting
from scratch, going right to the hardware counters and building from
there, is a big project--they've been hacking on oprofile for almost six
years now and still aren't suggesting it's release quality yet.

--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD

On 12/17/07, Decibel! <decibel@decibel.org> wrote:

Also, has anyone looked into adding a class of system calls that
would actually tell us if the kernel issued physical IO? I find it
hard to believe that other RDBMSes wouldn't like to have that info...

Non-blocking style interfaces can help here. On Windows, for instance,
a read returns data at the call site if it was satisfied by cache,
instead of invoking the asynchronous notification.

"Decibel!" <decibel@decibel.org> writes:

Also, has anyone looked into adding a class of system calls that would
actually tell us if the kernel issued physical IO? I find it hard to believe
that other RDBMSes wouldn't like to have that info...

Yeah, I think that's called "DTrace"

--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's 24x7 Postgres support!

On Dec 18, 2007, at 3:32 AM, Gregory Stark wrote:

Also, has anyone looked into adding a class of system calls that
would
actually tell us if the kernel issued physical IO? I find it hard
to believe
that other RDBMSes wouldn't like to have that info...

Yeah, I think that's called "DTrace"

That seems like way overkill...

When a read() call returns, surely the kernel knows whether it
actually issued a physical read request to satisfy that. I don't see
any reason why you couldn't have a version of read() that returns
that information. I also rather doubt that we're the only userland
software that would make use of that.
--
Decibel!, aka Jim C. Nasby, Database Architect decibel@decibel.org
Give your computer some brain candy! www.distributed.net Team #1828

"Decibel!" <decibel@decibel.org> writes:

When a read() call returns, surely the kernel knows whether it actually issued
a physical read request to satisfy that. I don't see any reason why you
couldn't have a version of read() that returns that information. I also rather
doubt that we're the only userland software that would make use of that.

I'm told this exists on Windows for the async interface. But AFAIK it doesn't
on Unix. The visibility into things like this is what makes DTrace so
remarkable.

I think there aren't many userland software interested in this. The only two
cases I can think of are databases -- which use direct I/O partly because of
this issue -- and real-time software like multimedia software -- which use
aren't so much interested in measuring it as forcing things to be preloaded
with stuff like posix_fadvise() or mlock().

I don't think DTrace is overkill either. The programmatic interface is
undocumented (but I've gotten Sun people to admit it exists -- I just have to
reverse engineer it from the existing code samples) but should be more or less
exactly what we need.

But the lowest-common-denominator of just timing read() and seeing if it took
long enough to involve either a context switch or sleeping on physical i/o
should be a pretty close approximation. The case where it would be least
accurate is when most or all of the data is actually in the cache. Then even
with a low false-positive rate detecting cache misses it'll still dominate the
true near-zero rate of cache misses.

We could mitigate that somewhat by describing it in the plan as something like

................... (... I/O fast=nnn slow=nnn)

instead of the more descriptive "physical" and "logical"

--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's 24x7 Postgres support!

Greg,

Gregory Stark wrote:

I don't think DTrace is overkill either. The programmatic interface is
undocumented (but I've gotten Sun people to admit it exists -- I just have to
reverse engineer it from the existing code samples) but should be more or less
exactly what we need.

You probably know this already. There are existing commands that use
the programmatic interface and would provide a good starting point. Here
are a couple:

http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/cmd/lockstat/lockstat.c
http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/cmd/dtrace/dtrace.c

One of my colleagues is in the process of putting a tutorial together
for how to do this, so if you decided to pursue this approach and need
assistance, please let me know.

Regards,
Robert