[GE users] Sort by sequence number question

Lönroth Erik erik.lonroth at scania.com
Thu Jul 19 15:07:45 BST 2007

    [ The following text is in the "iso-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

1. I'm using 6.0u8.

2. I completely re-installed one of my SGE cells and did only the minimal configuration required to make it possible to run a "sleep 60" job.

3. I set "sort by seq no"

bash-3.00$  qconf -ssconf | grep seq
queue_sort_method                 seqno

4. I configure the queues as follows:

bash-3.00$ qstat -F | egrep "seq|@"
master.101.q at ts101-1-0.sss.se. BIP   0/1       0.00     lx26-amd64    
short.101.q at ts101-1-1.sss.se.s BIP   0/4       0.00     lx26-amd64    
short.101.q at ts101-1-0.sss.se.s BIP   0/4       0.00     lx26-amd64    

5. I have "$fill_up" in my "PE" and as you see here, the queues are configured with the PE.

bash-3.00$ qconf -sq master.101.q | egrep "pe_list"
pe_list               make powerflow_ts101_pe
bash-3.00$ qconf -sq short.101.q | egrep "pe_list"
pe_list               make powerflow_ts101_pe

6. I manipulate the host I wan't to be considered last e.g. ts101-1-0:

bash-3.00$ qconf -sq short.101.q | egrep "hostlist|seq_no"
hostlist              ts101-1-0.sss.se.scania.com ts101-1-1.sss.se.scania.com
seq_no                0,[ts101-1-0.sss.se.scania.com=99]

bash-3.00$ qconf -sq master.101.q | egrep "hostlist|seq_no"
hostlist              ts101-1-0.sss.se.scania.com
seq_no                0

7. I submit the sleep-job, asking for 2 slots with that PE. It is dispatched, but not as I would think it would.

bash-3.00$ qstat -t
19 0.55500 slot-alloc sssler       r     07/19/2007 16:02:51 master.101.q at ts101-1-0.sss.se. MASTER
19 0.55500 slot-alloc sssler       r     07/19/2007 16:02:51 short.101.q at ts101-1-0.sss.se.s SLAVE

Apart from this - default configuration.


-----Original Message-----
From: Andreas.Haas at Sun.COM [mailto:Andreas.Haas at Sun.COM]
Sent: Wed 7/18/2007 4:31 PM
To: users at gridengine.sunsource.net
Subject: Re: [GE users] Sort by sequence number question
Hi Ravi,

On Wed, 18 Jul 2007, Ravi Chandra Nallan wrote:

> Hi Andreas,
> Andreas.Haas at Sun.COM wrote:
>> On Tue, 17 Jul 2007, Paul MacInnis wrote:
>>> The jobs here have no -soft options but they do have load thresholds:
>>> qname                 ser.q
>>> hostlist              @1g.hg @2g.hg @4g.hg
>>> seq_no                1965,[@2g.hg=2965],[@4g.hg=4965]
>>> load_thresholds       load_avg=1.5,mem_used=500M,[@2g.hg=load_avg=1.5, \
>>>                      mem_used=1.5G],[@4g.hg=load_avg=1.5,mem_used=3.5G]
>>> suspend_thresholds    NONE
>>> qname                 bg.q
>>> hostlist              @2g.hg @4g.hg
>>> seq_no                2969,[@4g.hg=4969]
>>> load_thresholds       load_avg=1.5,mem_used=1.5G,[@4g.hg=load_avg=1.5, \
>>>                      mem_used=3.5G]
>>> suspend_thresholds    load_avg=2.5
>>> Each 2G and 4G node has a ser.q and a bg.q queue instance, each with same
>>> load_thresholds, but scheduler has a definite preference for the bg.q
>>> instance, inspite of the higher seqno!  Perhaps in time an explanation
>>> will appear ...
>> Could you try whether behaviour changes anyhow when you set load_thresholds
>> to NONE with both queues? Just temporarily for testing purposes. Load 
>> thresholds make setups always hard to survey, whereas setups without load 
>> thrsholds are fairly deterministic.
>> Andreas
> But does load_threshold play a role in choosing the queue when the 
> queue_sort_method is set seqno?

Sure it does.

> And if load_threshold were to make a queue unusable, wouldn't the q be set to 
> alarm state?


> Also I noticed that Erik was able to reproduce the prob with pe jobs. I am 
> not sure if the array jobs had a similar problem.

To me these are two different problems: (a) With Eriks parallel job allocation 
problem I already fail to understand why it behaved the way as he described it 
for 5.3 and how he needs it, whereas (b) Pauls sequential job problem looks to me 
like the outcome confusing load thresholds setup.

> I couldn't reproduce it with neither pe not array jobs, must be some setting 
> that is effecting it.

Yep. At least Pauls basic setup worked in my cluster here. For that
reason I asked Paul to disable load thresholds temporarily.


To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net

    [ Part 2, "ATT9730812.txt"  Text/PLAIN (Name: "ATT9730812.txt") ~213 ]
    [ bytes. ]
    [ Unable to print this part. ]

More information about the gridengine-users mailing list