You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@daffodil.apache.org by "Mike Beckerle (Jira)" <ji...@apache.org> on 2021/01/21 20:59:00 UTC
[jira] [Closed] (DAFFODIL-1823) Improve performance report with
kB/sec and speedup factor
[ https://issues.apache.org/jira/browse/DAFFODIL-1823?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Mike Beckerle closed DAFFODIL-1823.
-----------------------------------
Resolution: Not A Problem
Performance report is not part of Daffodil
> Improve performance report with kB/sec and speedup factor
> ---------------------------------------------------------
>
> Key: DAFFODIL-1823
> URL: https://issues.apache.org/jira/browse/DAFFODIL-1823
> Project: Daffodil
> Issue Type: Improvement
> Components: CLI, Infrastructure, TDML Runner
> Reporter: Mike Beckerle
> Assignee: Dave Thompson
> Priority: Major
>
> Our performance reports currently are good at informing us about regrressions.
> They are not useful to someone trying to gauge whether Daffodil is fast enough to meet their needs.
> The simplest initial fix for this is to provide a KBytes/sec statistic. For parsing this is the rate at which input data is consumed. For unparsing it is the rate at which output data is emitted.
> (I.e., I/O related to the infoset in the middle isn't relevant).
> That additional measure, which is just (files/sec) / (kbytes/file) would be a great improvement to this report.
> As a spreadsheet of data, I would prefer to see the things mashed together into test names spread out into columns:
> E.g., army-drrs_1t_1.1m_60
> as {code}
> {{name nThreads file size file size units nFiles
> army-drrs 1 1.1 m 60 }}
> {code}
> For file-size I'd prefer to see it standardized to kbytes actually rather than having this separate units column.
> For multi-threaded tests, a statistic to add is speedup factor. This is the Nthread speed (kbytes/sec) divided by the 1 thread kbytes/second, divided by N (number of threads).
> Perfect scaling up has speedup factor of 1.
> Most things will have speedup factors < 0.5
> As you increase the number of threads, the speedup factor should level off and eventually drop.
> The above stats would give users who understand the approximate message size of the data they are handling, to get an intuition for "messages/sec" and the speedup curve they should expect from multi-threading.
> E.g., for Link16, a typical message is about 32 bytes long.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)