Thread: optimizing a cpu-heavy query
Folks,
I'm trying to optimize the following query that performs KL Divergence [1]. As you can see the distance function
operateson vectors of 150 floats.
The query takes 12 minutes to run on an idle (apart from pgsql) EC2 m1 large instance with 2 million documents in the
docstable. The CPU is pegged at 100% during this time. I need to be able to both process concurrent distance queries
andotherwise use the database.
I have the option of moving this distance calculation off of PG but are there other options?
Is there anything clearly wrong that I'm doing here?
Would it speed things up to make the float array a custom data type backed by C code?
Thanks in advance, Joel
[1] http://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence
---
CREATE DOMAIN topics AS float[150];
CREATE DOMAIN doc_id AS varchar(64);
CREATE TABLE docs
(
id serial,
doc_id doc_id NOT NULL PRIMARY KEY,
topics topics NOT NULL
);
CREATE OR REPLACE FUNCTION docs_within_distance(vec topics, threshold float)
RETURNS TABLE(id doc_id, distance float) AS $$
BEGIN
RETURN QUERY
SELECT *
FROM (SELECT doc_id, (SELECT sum(vec[i] * ln(vec[i] / topics[i]))
FROM generate_subscripts(topics, 1) AS i
WHERE topics[i] > 0) AS distance
FROM docs) AS tab
WHERE tab.distance <= threshold;
END;
$$ LANGUAGE plpgsql;
--------------------------------------------------------------------------
- for hire: mac osx device driver ninja, kernel extensions and usb drivers
---------------------+------------+---------------------------------------
http://wagerlabs.com | @wagerlabs | http://www.linkedin.com/in/joelreymont
---------------------+------------+---------------------------------------
Joel Reymont <joelr1@gmail.com> writes:
> I'm trying to optimize the following query that performs KL Divergence [1]. As you can see the distance function
operateson vectors of 150 floats.
> CREATE OR REPLACE FUNCTION docs_within_distance(vec topics, threshold float)
> RETURNS TABLE(id doc_id, distance float) AS $$
> BEGIN
> RETURN QUERY
> SELECT *
> FROM (SELECT doc_id, (SELECT sum(vec[i] * ln(vec[i] / topics[i]))
> FROM generate_subscripts(topics, 1) AS i
> WHERE topics[i] > 0) AS distance
> FROM docs) AS tab
> WHERE tab.distance <= threshold;
> END;
> $$ LANGUAGE plpgsql;
Yikes. That sub-select is a mighty expensive way to compute the scalar
product. Push it into a sub-function that takes the two arrays and
iterates over them with a for-loop. For another couple orders of
magnitude, convert the sub-function to C code. (I don't think you need
a whole data type, just a function that does the scalar product.)
regards, tom lane
Tom, On Apr 26, 2011, at 5:00 PM, Tom Lane wrote: > For another couple orders of magnitude, convert the sub-function to C code. (I don't think you need > a whole data type, just a function that does the scalar product.) That's a 30x speedup, from 12 minutes down to 38s. Thanks Tom! -------------------------------------------------------------------------- - for hire: mac osx device driver ninja, kernel extensions and usb drivers ---------------------+------------+--------------------------------------- http://wagerlabs.com | @wagerlabs | http://www.linkedin.com/in/joelreymont ---------------------+------------+---------------------------------------
Joel Reymont <joelr1@gmail.com> writes:
> On Apr 26, 2011, at 5:00 PM, Tom Lane wrote:
>> For another couple orders of magnitude, convert the sub-function to C code. (I don't think you need
>> a whole data type, just a function that does the scalar product.)
> That's a 30x speedup, from 12 minutes down to 38s. Thanks Tom!
Huh, I would've bet on a lot more actually. The nodeFunctionscan and
nodeAgg code must not be as inefficient as it looks on the surface ...
regards, tom lane
2011/4/27 Tom Lane <tgl@sss.pgh.pa.us>: > Joel Reymont <joelr1@gmail.com> writes: >> On Apr 26, 2011, at 5:00 PM, Tom Lane wrote: >>> For another couple orders of magnitude, convert the sub-function to C code. (I don't think you need >>> a whole data type, just a function that does the scalar product.) > >> That's a 30x speedup, from 12 minutes down to 38s. Thanks Tom! > > Huh, I would've bet on a lot more actually. The nodeFunctionscan and > nodeAgg code must not be as inefficient as it looks on the surface ... Did you mean in that case you can optimize it by collapsing those nodes into one? Regards, -- Hitoshi Harada
Hitoshi Harada <umi.tanuki@gmail.com> writes:
> 2011/4/27 Tom Lane <tgl@sss.pgh.pa.us>:
>> Joel Reymont <joelr1@gmail.com> writes:
>>> That's a 30x speedup, from 12 minutes down to 38s. Thanks Tom!
>> Huh, I would've bet on a lot more actually. �The nodeFunctionscan and
>> nodeAgg code must not be as inefficient as it looks on the surface ...
> Did you mean in that case you can optimize it by collapsing those
> nodes into one?
No, just that the overhead of that code looks like it ought to be much
more than the cost of a few multiplications ...
regards, tom lane