[ 
https://issues.apache.org/jira/browse/TS-54?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13876305#comment-13876305
 ] 

Leif Hedstrom commented on TS-54:
---------------------------------

[~jplevyak] Are there still actionable items on this bug? Moving out to 
"sometime" for now. Also, [~psudaemon] This looks like something in the area 
where you had some thoughts an ideas as well?

Also, take a look TS-2153, where we're considering changing events for 
timeouts, and instead moving to LRUs. It not only is more efficient (we think), 
by eliminating a lot of events, we feel it's a lot easier to manage from an 
operations perspective. It's virtually impossible today to scale the number of 
connections you allow with the 6 or so timeout configurations.

> UnixNet cleanup, encapsulation of event subsystem
> -------------------------------------------------
>
>                 Key: TS-54
>                 URL: https://issues.apache.org/jira/browse/TS-54
>             Project: Traffic Server
>          Issue Type: Improvement
>          Components: Cleanup
>         Environment: all unixesque
>            Reporter: John Plevyak
>            Assignee: John Plevyak
>             Fix For: sometime
>
>         Attachments: proposed-jp-v1-List.h, ts-List-net-cleanup-jp-v2.patch, 
> ts-List_and_net-cleanup-jp-v1.patch
>
>   Original Estimate: 72h
>  Remaining Estimate: 72h
>
> The UnixNet subsystem was modified for epoll, but lots of the old data 
> structures, data members and code
> remain for the old "bucket" approach.
> The epoll code should also be encapsulated to simplify support for other 
> platforms and a possible move
> to an event library.
> The current code is complicated by limitations in Queue which require 
> specifying the link field for every
> use, but which can be fixed by in the template.
> Finally, the current code does an unnecessary allocation for the epoll struct 
> which should be part of the NetVConnection etc.
> and it takes a lock for the enable_queue which can be avoided by using the 
> non-locking AtomicSSL.
> This work is also good preparation for evaluating libev or libevent as it 
> will reduce the amount of code which
> will have to be changed.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)

Reply via email to