Task duration correction factor

[AF>EDLS]zOU
[AF>EDLS]zOU
Joined: 5 May 15
Posts: 53
Credit: 299,075,257
RAC: 524,244
Topic 228570

Can anyone explain in layman's terms the "Task duration correction factor" value ?

 

Arc A750 system: Task duration correction factor:0.180861

Dual GTX980 system: Task duration correction factor:0.355147

Pure CPU system: Task duration correction factor:1.591853

 

 

[AF>EDLS]zOU
[AF>EDLS]zOU
Joined: 5 May 15
Posts: 53
Credit: 299,075,257
RAC: 524,244

ok,   I found

ok,

 

I found this:

An TDCF less than 1 means the host is faster than the estimate and greater than one is slower.

 

so that helps the project sending enough WU so they can be completed before the deadline, if I understand it correctly.

Sending too many WU would get some WU aborted, sending not enough would force the host to ask for WU too often

mikey
mikey
Joined: 22 Jan 05
Posts: 9,834
Credit: 1,349,141,526
RAC: 860,086

[AF>EDLS

[AF>EDLS wrote:

zOU]

ok,

 

I found this:

An TDCF less than 1 means the host is faster than the estimate and greater than one is slower.

 

so that helps the project sending enough WU so they can be completed before the deadline, if I understand it correctly.

Sending too many WU would get some WU aborted, sending not enough would force the host to ask for WU too often

Yes that's correct but it changes over time as you work on different kinds of tasks or the project changes a type of task so it runs shorter or longer, it also takes time to settle down into a fairly constant number

archae86
archae86
Joined: 6 Dec 05
Posts: 3,116
Credit: 6,398,522,367
RAC: 3,143,868

mikey wrote: it also takes

mikey wrote:

it also takes time to settle down into a fairly constant number

A particular gripe with the system is that it NEVER settles down to a fairly constant number in many cases where you have more than one type of Einstein task running on a system.  The "correct" DCF for that system for one type of task differs from that for another type of task--sometimes by an order of magnitude.  So the current value of DCF depends heavily on which tasks have completed recently.  This in turn leads to surges in work fetch activity, fluctuations in work queue size, and sometimes to "panic mode" prioritization of tasks deemed in danger of failing to meet their deadlines.

I dislike this enough that I have usually run Einstein with just one type of task allowed by my preferences at any given time.

As it happens, in recent weeks I've made an exception to this rule, and am currently running two types: MeerKAT GPU tasks and GRP #1 GPU tasks.  These differ by a small enough amount, and I have my queue request set low enough, that the problems I've mentioned have not been bothering me.

 

mikey
mikey
Joined: 22 Jan 05
Posts: 9,834
Credit: 1,349,141,526
RAC: 860,086

archae86 wrote:mikey

archae86 wrote:

mikey wrote:

it also takes time to settle down into a fairly constant number

A particular gripe with the system is that it NEVER settles down to a fairly constant number in many cases where you have more than one type of Einstein task running on a system.  The "correct" DCF for that system for one type of task differs from that for another type of task--sometimes by an order of magnitude.  So the current value of DCF depends heavily on which tasks have completed recently.  This in turn leads to surges in work fetch activity, fluctuations in work queue size, and sometimes to "panic mode" prioritization of tasks deemed in danger of failing to meet their deadlines.

I dislike this enough that I have usually run Einstein with just one type of task allowed by my preferences at any given time.

As it happens, in recent weeks I've made an exception to this rule, and am currently running two types: MeerKAT GPU tasks and GRP #1 GPU tasks.  These differ by a small enough amount, and I have my queue request set low enough, that the problems I've mentioned have not been bothering me. 

Add in if you run both cpu and gpu tasks on the same pc and it's always in a state of flux, you seem to have found a sweet spot though

That's why I have so many pc's, it means I can run cpu tasks from Einstein on this pc and gpu tasks on that pc and the cache is going crazy giving me 200 cpu tasks and then telling me 'the cache is full' when I run out of gpu tasks!!

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.