libcloud-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Russell Alexander Spitzer (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (LIBCLOUD-527) GCE Libcloud Driver list_nodes and list_volumes Scaling Improvement
Date Fri, 07 Mar 2014 20:42:42 GMT

    [ https://issues.apache.org/jira/browse/LIBCLOUD-527?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13924312#comment-13924312
] 

Russell Alexander Spitzer commented on LIBCLOUD-527:
----------------------------------------------------

Np :) I didn't notice it until I called list_nodes when I had 100 instances up. 

My solution in the branch above is to remove the setting of extra['boot_volume'] from to_node
and move it into list_nodes, and ex_get_node. In get node we do an additional call for that
volume's info and attach it so the behavior is the same. For list_nodes we call list_volumes
and match each boot volume from that request with it's instance.

For list_volumes I noticed that we actually have the zone info before calling _to_storage_volume
so I just apply the zone info without the additional request. When no zone is supplied we
make one request for each zone present as opposed to once per volume. 

> GCE Libcloud Driver list_nodes and list_volumes Scaling Improvement
> -------------------------------------------------------------------
>
>                 Key: LIBCLOUD-527
>                 URL: https://issues.apache.org/jira/browse/LIBCLOUD-527
>             Project: Libcloud
>          Issue Type: Improvement
>    Affects Versions: 0.14.0-beta3
>            Reporter: Russell Alexander Spitzer
>
> The current code for the GCE Driver has two significant bottle necks in the list_nodes
and list_volumes code. In both cases the slowdown can be traced to the response to object
methods _to_node and _to_storage_volume. These methods are called in serial from both list_node
and list_volume methods and each call to one of these methods requires an additional http
request. For list_nodes the price is doubled since each _to_node request requires a _to_storage_volume
request.
> Basically, list_nodes and list_volumes require 1 + 2 * Nodes and 1 + Volumes requests
to complete. I've found that at large numbers of nodes this is a considerable pain point.
> I've been working on a branch at
> https://github.com/RussellSpitzer/libcloud/tree/GCE_perf_to_node
> that I would be happy to contribute upstream if there is any desire for it.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message