jeffrey.kissel@LIGO.ORG - posted 13:32, Monday 21 August 2023 (72356)
HAMs 3 thru 7, BSC 123 SUS Computer Clock Cycle Usage Thus Far during O4 :: All User Model SUS that are to be on-the-move below Run Below 25% Usage
J. Kissel As I continue game-planning the integration of O5 suspensions into the existing system, (see progress thus far in G2301306) one of the questions is the percentage of computer usage needed on a given computer core running different numbers and amounts of suspensions. Given the sheer amount of new SUS coming in coupled with the limited number of computers and IO chassis, we'll need to pull a similar computer core assignment re-arrangement game that we've done before previous observing runs -- (for O4 see G2101636, and implementation aLOGs LHO:59651 for HAM56 and LHO:59669 for HAM7. For O3 see LHO:38827). Summary of findings shown in the attached trends of the CPU_METER_MAX values for SUSB123 and SUSH34 and SUSH56 and SUSH7 since May 24 2023 16:00 UTC: :: All 65 kHz IOP Models, be it SUSH34, SUSB123, SUSH56, or SUSH7 are consistently using 8 usec / 15 usec ~ 53% of capacity. On SUSB123, :: The ITM 16 kHz user models consume the most clock cycle time, running with maximums of ~26 usec / 61 usec ~42%. (I presume the abnormally large computation comes from violin mode damping) :: The BS 16 kHz user model only consumes 14 / 61 ~ 23% of its core. On SUSH34, :: All three HSTSs, MC2, PR2, and SR2 run between 13 / 61 usec ~ 21% of core capacity. MC2 had "run the longest" at 14 and I presume that's from the fast longitudinal global control filtering (where PR2 and SR2 only receive slow angular drift control), but the restart after reboot from the dolphin crash on July 30 2023 (see LHO:71829) seems to have brought it back in symmetry with the other two at 13. On SUSH56 and SUSH7, :: all SUSH56 and SUSH7 16 kHz user models are using at most only 15 / 61 usec ~ 25% (it's SUSSRM running one globally controlled HSTS, and SUSSQZIN running three ZMs [two HPDS and one HDDS] and the OPO). Some, like SUSSQZOUT which are only running damping loops on ZM4, ZM5, and ZM6 top masses (two HPDSs and one HSDS, respectively) are using as little as 8 / 61 ~13% Their 65 kHz IOPs are running at ~50% capacity. :: The SUSAUXH7 model is running at 4 kHz, and consumes only 4 / 244 ~ 6% of its core on SUSH7. I'll note that no one has populated the coil driver monitor filter banks with any filters, but I'm sure even if it was chock full I doubt it would even come close to "breaking the bank." This bodes well for re-organizing models for O5; we're not computationally limited in doing so for any suspensions on the move, even if we keep running things at 16 kHz. Further, we can consider combining even more suspensions' controls into single models if we expect the SUS's filter computation to be light [e.g. only damping like in SUSSQZOUT].
Images attached to this report
