[pgrouting-users] "invalid memory alloc request size" on large graphs

Phyks phyks+pgrouting at phyks.me
Sat Jan 21 13:12:54 PST 2023


Hi,

> 1.) set the highest "work_mem" for the session and close all other running
> programs ..
> 
> SET work_mem = '8GB';
> SELECT pgr_createTopology( ... )
> -- set back for the default value;
> SET work_mem = '1GB';

Same issue :

=# SET work_mem = '8GB';
SET
=# (SELECT component, node FROM pgr_connectedComponents(
         'SELECT id, source, target, cost, cost AS reverse_cost FROM edges'
));
ERREUR:  invalid memory alloc request size 1080000000
CONTEXTE : fonction SQL « pgr_connectedcomponents », instruction 1


> 2.) Linux:  Enable Swap space (  ~ 20GB  or  2x of your RAM )
> https://www.digitalocean.com/community/tutorials/how-to-add-swap-space-on-ubuntu-20-04
> 
> 3.) optimize table variables storage spaces ( to less space )
>      - change your bigint variables to integer
>      for (about 40M edges) an integer  [  -2147483648 to +2147483647 ]
> should enought
>      https://www.postgresql.org/docs/current/datatype-numeric.html

As far as I understand, `pgr_connectedComponents` as run in the example 
above should only query the `edges` table. I already have `source` and 
`target` set as regular `integer` there.

I have the impression the malloc is failing for another reason that OOM:
1/ Malloc size should be about 1GB as per the error message. This is far 
less than my free RAM and seems to be a fixed buffer size somewhere.
2/ I have only 2GB RAM used on my system (hence 10GB free RAM) and do 
not see any increase of RAM usage between the start of the 
`pgr_connectedComponents` query and the error output.

For these reasons, I do not expect swap / RAM changes to have any effect 
here, but instead a configuration option in PostgreSQL / pgRouting to 
allow allocations of more than 1GB?

It seems I'm hitting https://postgrespro.com/list/thread-id/1549721 
which references https://github.com/pgRouting/pgrouting/issues/291 but I 
don't see much solutions there apart from using bounding boxes (which is 
easily done for routing, much less for contraction / connected 
components analysis).

Thanks!
Phyks

> Phyks <phyks+pgrouting at phyks.me> ezt írta (időpont: 2023. jan. 21., Szo,
> 14:06):
> 
>> Hi,
>>
>> I'm trying to use pgRouting on a quite large graph (about 40M edges). My
>> graph structure is expected to be quite simple and not very connected
>> (being mostly a simple tree structure).
>>
>> I came across the discussion on speeding up `pgr_createTopology` for
>> large networks
>> (
>> https://github.com/pgRouting/pgrouting/discussions/2255#discussioncomment-4739376)
>>
>> and am trying to put it in practice for my use case. Moving from
>> `pgr_createTopology` to `pgr_extractVertices` gives a huge increase in
>> processing time and memory requirements.
>>
>> However, when trying to run `pgr_*` (`pgr_dijkstra`, `pgr_contraction`
>> or `pgr_connectedComponents` for example) on the subsequent graph, I am
>> facing a malloc issue : invalid memory alloc request size 1080000000.
>>
>> This is easily worked around with `pgr_dijkstra` since routing is mainly
>> local and the edges table can easily be filtered with a bounding box.
>> This is not as easy for `pgr_contraction` or `pgr_connectedComponents`
>> where tiling the computation (through bbox) requires some extra
>> post-processing.
>>
>> I'm not sure about the units of the `malloc` error message, but it seems
>> to me this is about 1GB allocation which should be no issue on the
>> machine and setup I am running. For the record, my postgresql is having
>> either a basic configuration (out of the box install) or a pgtune-d one,
>> same issue in both cases. My machine has a bit more than 10GB of free RAM.
>>
>> Would anyone have more insights on this issue which could help me better
>> understand and work around it? This is particularly frustrating since
>> computations on a quarter of the table (about 10M edges) work perfectly
>> fine and in a very reasonable amount of time.
>>
>> Thank you in advance,
>> Best,
>> _______________________________________________
>> Pgrouting-users mailing list
>> Pgrouting-users at lists.osgeo.org
>> https://lists.osgeo.org/mailman/listinfo/pgrouting-users
>>
> 
> 
> _______________________________________________
> Pgrouting-users mailing list
> Pgrouting-users at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/pgrouting-users


More information about the Pgrouting-users mailing list