Reports until 10:10, Tuesday 15 August 2017
H1 SUS (CDS, DetChar, ISC, SUS)
jeffrey.kissel@LIGO.ORG - posted 10:10, Tuesday 15 August 2017 - last comment - 16:53, Tuesday 12 September 2017(38198)
H1 QUAD CPU Cycle Turn Around Time during O2 
J. Kissel, D. Barker

Posting some data to support a future ECR to add more violin mode banks to the QUADs, here's the CPU Cycle Turn Around Time during O2 for the front-end models that run the QUADs. The ITMs have been running around 42-45 [usec], and the ETMs have been running around 32-33 [usec] during O2 (see first attachment).

Remember, two things have happened that have *reduced* the CPU turn-around time over the past few years (see second attachment):
- the end-station front-end computers were upgrade to better CPUs back in Feb 2016 (see LHO aLOG 25474); prior to that they'd typically run at around ~55 [usec].
- All QUAD model's data storage list was pruned in Oct 2016 (see LHO aLOG 30821). This knocked about 10-15 usec off of the ITM time, and a further ~2-3 [sec] off of the ITMs.

In short: there should be plenty of CPU time to add oodles of violin mode damping filters.
Images attached to this report
Comments related to this report
david.barker@LIGO.ORG - 10:15, Tuesday 15 August 2017 (38199)
Link

and of course we could install the faster computer for h1susb123

jeffrey.kissel@LIGO.ORG - 16:53, Tuesday 12 September 2017 (38619)CDS
Link 

D. Barker, J. Kissel

I (and Norna, and Dennis) asked Dave if we could better quantify "there should be plenty of CPU time to add oodles of violin mode damping filters."

As such, he says (via email) 
   On Sep 6, 2017, at 4:39 PM, David Barker  wrote:  
   [...]

   For my first set of tests I took the ITMX_L2_DAMP_MODE10 [, an example, 
   fully loaded, in-use violin] filter module and duplicated it 32 times 
   on two test models: x1susdactest and x1susfiltertest.
   [piping junk data into the input of the bank to make sure the filters
    were computing on something]

   x1susdactest was already running of the fast computer x1susex (doing my 
   18bit-DAC testing). I then made a copy of this model for x1susfiltertest, 
   which ran on the slow computer x1sush34.

   x1susex  = Intel Xeon E5-2690 v2 @ 3.00GHz (fast 10-core computer)
   x1sush34 = Intel Xeon X5680 @ 3.33GHz (slow 6-core computer)

   cpu		no filters		32 filters
   slow		3uS			7uS
   fast		4uS			7uS

   not much in it for 32 filter modules, about 1uS increase for every 10 FMs loaded.
   [...]

So, extrapolating Dave's data, if we want to increase from 10 to 40 violin mode filters, that would be adding 3 [us] to the clock cycle turn around time.

As Dave mentions, we'll likely switch h1susb123 over to the faster computer type he mentions, which means the ITMs (which had high turn-around at 42-45 [us]) will run more like the ETMs did during the tail end of O2, at ~32-33 [us].

That means for ETMs and ITMs, with the inclusion of these extra filters, the turn-around time would likely only increase to 36-37 [us], indeed still with plenty of head room against the limit of 61 [us].