@IrisHadarStltd ,
Note, the instances used and saturation reported by ArcGIS Monitor are approximation of Instances in Use, not Instances Running (which is measured by arcsoc.exe process), see below:
Also, the response time degradation you showed is present during time where there are 11 instances (the right side of the chart, after 10). So this degradation does not correlate well with the arcsoc.exe process.
This performance fluctuation could be due to different request/response, e.g. different size of the features fetched and not instance saturation. To verify it, you'd need to set logs temporarily to verbose and verify elapsed time and features returned.
One potential explanation of your case of higher instances running during this time and low instances in use might be the result of the previous spike (e.g. 8:00) and these idle instances are waiting to be recycled down. This is controlled by The maximum time an idle instance can be kept running (default 1800 sec), see https://enterprise.arcgis.com/en/server/latest/publish-services/windows/tuning-and-configuring-servi...
To diagnose it, please provide request rate, request received and request failed for this time and 1 hour before. Also, any chance min max setting where changed around that time? Finally, please take a clear screenshot (the images you shared difficult to read) or if easier, please email Excel Report to me?