Re: Softcrash watchdog

Hey fullzero, i have a question,

without a doubt my biggest problem right now is that when my miner crashes it takes the whole rig down with it, everything gets stuck, SSH barely works, average system load jumps to 14.5!! and Xorg takes up 100% of the CPU, its so bad that none of the standard reboot commands work, they just do nothing, the only thing that actually reboots the rig in this state is "echo b > /proc/sysrq-trigger" so i've set up a script that checks the average system load and if its over 2 it uses the command to reboot, and it works, but i dont like this "solution", yesterday after a reboot nvOC got corrupted somehow, lost my customized oneBash and the whole system became read-only (thankfully i had a oneBash backup that was only a few days behind).

so the question is, what can i do to relive this Xorg error, i run a 7 card rig and never plan on going for a higher number, what can i do with Xorg that would fix this?

Thanks.

@ tempgoga

It seems that whenever a soft crash occurs most of the cards drop to zero, so while the display/keyboard is unresponsive you can catch the soft crash from nvidia-smi. The script below checks card utilization, if it drops below 90% it counts down a minute and if mining hasn't resumed it reboots the system.
This seems to have worked at least once in my case (only got one soft crash this weekend) and the system recovered as expected.
the threshold values work for my setup but others may find different values optimal

Also if anyone knows a way to iterate the if && statements we can get the card count from "cards=$(nvidia-smi -L | wc -l); echo $cards" but the way below also works with manual editing to adjust the watchdog for the number of cards in you individual system.
___________

#!/bin/bash
#m1
threshold=90
while sleep 5
do number=$(nvidia-smi |grep % |awk '{print $13}' |tr -d %)
set -- $number
echo -e "$@"
# The "if and" statements below need to be manually adjusted to match the number of cards in your system
# If you have 5 cards, leave is as, if a different number of cards remove or add the && statements as needed as in the example below
   if [[ "$1" -gt "$threshold" ]] && \
   [[ "$2" -gt "$threshold" ]] && \
   [[ "$3" -gt "$threshold" ]] && \
   [[ "$4" -gt "$threshold" ]] && \
   [[ "$5" -gt "$threshold" ]]
# && \
# [[ "$6" -gt "$threshold" ]]
   then i=12
   echo OK
   else echo $((i--))
   fi
   if [ $i -le 0 ]
   then echo $(date) REBOOT due to soft crash >>~/watchdog.log
   sleep -5
   sudo shutdown now -r
   fi
done
___________

Hey thats funny I just made a script doing something similar, although it checks the powerdraw.
Here it is:

Code:

#!/bin/bash

# Miner restart script V001
# By Maxximus007
# for nvOC by fullzero
#
# POWERLIMIT MUST BE SET IN oneBash

#########################
### BELOW CODE, NO NEED FOR EDITING
#########################
echo "$(date) - Starting miner restart script." | tee -a ${LOG_FILE}
# Creating a log file to record restarts
LOG_FILE="/home/m1/restartlog.txt"
if [ ! -e "$LOG_FILE" ] ; then
   touch "$LOG_FILE"
fi

while true
do
sleep 60

GPUS=$(nvidia-smi --query-gpu=count --format=csv,noheader,nounits | tail -1)

gpu=0
COUNT_LOW_POWER=0

while [ $gpu -lt $GPUS ]
do
  { IFS=', ' read POWERDRAW POWERLIMIT; } < <( nvidia-smi -i $gpu --query-gpu=power.draw,power.limit --format=csv,noheader,nounits)

  let POWER_DIFF=$( printf "%.0f" $POWERLIMIT )-$( printf "%.0f" $POWERDRAW )

  # If current draw is 30 Watt lower than the limit count them:
  if [ "$POWER_DIFF" -gt "30" ]
  then
   let COUNT_LOW_POWER=COUNT_LOW_POWER+1
  fi

  let gpu=gpu+1
done

if [ $COUNT_LOW_POWER -eq $GPUS ]
then
  echo "$(date) - Power draw is too low: kill miner and oneBash" | tee -a ${LOG_FILE}
  # If miner runs in screen 'miner' kill the screen
  screen -X -S miner kill
  # Best to restart oneBash - settings might be adjusted already
  kill ps -ef | awk '$NF~"oneBash" {print $2}'
else
  echo "$(date) - All good! Will check again in 60 seconds"
fi

done

You can combine the above with your code, and find the utilization like this:

Code:

nvidia-smi -i 1 --query-gpu=utilization.gpu --format=csv,noheader,nounits

You have to iterate the GPU, starting at 0 to get them all

Okay I've combined the two, perhaps this will work for most of us:

Code:

#!/bin/bash

# Miner restart script V002
# By Maxximus007 && IAmNotAJeep
# for nvOC by fullzero
#

#########################
### BELOW CODE, NO NEED FOR EDITING
#########################
echo "$(date) - Starting miner restart script." | tee -a ${LOG_FILE}
# Creating a log file to record restarts
LOG_FILE="/home/m1/restartlog.txt"
if [ ! -e "$LOG_FILE" ] ; then
   touch "$LOG_FILE"
fi

MIN_UTIL=90
RESTART=0

while true
do
sleep 60

GPUS=$(nvidia-smi --query-gpu=count --format=csv,noheader,nounits | tail -1)

gpu=0
COUNT=0

while [ $gpu -lt $GPUS ]
do
  { IFS=', ' read UTIL; } < <( nvidia-smi -i $gpu --query-gpu=utilization.gpu --format=csv,noheader,nounits)

  let UTILIZATION=$( printf "%.0f" $UTIL )

  # If current utilizations lower than the limit count them:
  if [ $UTILIZATION -lt $MIN_UTIL ]
  then
   let COUNT=COUNT+1
  fi

  let gpu=gpu+1
done

if [ $COUNT -eq $GPUS ]
then
  if [ $RESTART -gt 1 ]
  then
   echo "$(date) - Utilization is too low: reviving did not work so restarting system" | tee -a ${LOG_FILE}
   sudo shutdown now -r
  fi
  echo "$(date) - Utilization is too low: kill miner and oneBash" | tee -a ${LOG_FILE}
  # If miner runs in screen 'miner' kill the screen
  screen -X -S miner kill
  # Best to restart oneBash - settings might be adjusted already
  kill ps -ef | awk '$NF~"oneBash" {print $2}'
  let RESTART=RESTART+1
else
  echo "$(date) - All good! Will check again in 60 seconds"
fi

done

Pretty cool! I'll try it tonight, lets hope this put the softcrash issues behind us.

I will try this out as well; good work.

@ Maxximus007
Thanks for putting these together, great collab!
I'm not a bash expert, so maybe I'm reading this wrong, but here are some thoughts.
The combined code seems to be evaluating each gpu individually for the fault condition to be met, which means if one fails and you have say 5 other cards working then it keeps going until all the cards give reduced output since all of them have to fail individually to increment the counter?So if 5/6 fail we keep going? (Again just looking at it and tracing it in my head so maybe I'm reading wrong).
The way I was thinking about it, is that I wanted all the cards to work at above 90% efficiency and reboot as soon as any card strays beyond the threshold - this is why I did the "if and" statement and didn't iterate though "if" statements alone (I didn't know how to iterate "if and" based on an unknown number of cards lol). I had a version giving 6xOK and such but I think it's more efficient to just get 1xOK if ALL meet the 90% criteria and start the countdown as soon as anything is out of norm - and if the miner recovers, flush the counter. I observed a number of these conditions with Claymore where it recovers half the time, but then eventually craps out and the script kicks in. I haven't seen it on my Genoil rig yet since my other script has kept it in check without any softcrash for day 3 now.

A thought about the power draw as threshold measure - it is power limit/card specific and I guess people would need to tune their power threshold to their power limit so I agree it's best to use gpu util. (My cards are at 82W limit for example).
Thoughts?

The code checks each cards individually, at times (with Claymore, not Genoil) I've seen that Util (or Powerpraw) is dropping, maybe even below 90 for a few seconds. In order not to generate too much restarts I check all cards. We can lower this or make it so that each of us can decide when it should reboot.
I've combined the restart/reboot so that the first attempt is to restart miner. If that doesn't work, we reboot the machine. We might want to reset the reboot counter after a while, so we don't loose time with a full reboot.

In the first code I checked Powerdraw -> if 30 Watt less than Powerlimit there might be something wrong. Idling cards use around 10 Watt, so that works for all I think. We can combine this with Util if that helps.

So sure we can make it more advanced, we just have to determine the right parameters. Hope others can let us know in what circumstances they see hanging miners. Just one card, or more or everything? Is Util back to zero? or hanging on to 100%?