mersenneforum.org GPU to 72 status...
 Register FAQ Search Today's Posts Mark Forums Read

 2014-11-19, 04:31 #3158 kladner     "Kieren" Jul 2011 In My Own Galaxy! 2×4,957 Posts Just sayin': two factors since I got into the DC catch-up project.
 2014-11-19, 14:13 #3159 Chuck     May 2011 Orange Park, FL 839 Posts Is the procedure to send out DC work when needed to people selecting "Let GPU72 Decide" in effect?
2014-11-19, 15:04   #3160
Mark Rose

"/X\(‘-‘)/X\"
Jan 2013
Ͳօɾօղէօ

3·72·19 Posts

Quote:
 Originally Posted by Chuck Is the procedure to send out DC work when needed to people selecting "Let GPU72 Decide" in effect?
It's probably not needed at the moment, given all the people manually working on DCTF. We still need work happening on TF.

2014-11-19, 15:53   #3161
chalsall
If I May

"Chris Halsall"
Sep 2002

100011010100002 Posts

Quote:
 Originally Posted by Chuck Is the procedure to send out DC work when needed to people selecting "Let GPU72 Decide" in effect?
It isn't, but as Mark said, it's not needed.

2014-11-19, 16:55   #3162
chalsall
If I May

"Chris Halsall"
Sep 2002

24×5×113 Posts

Quote:
 Originally Posted by Mark Rose The cg1.4xlarge is only available in us-east-1 and eu-west-1. Of those, only in one available zone in us-east-1 has the spot price been hovering around $0.135. Mark: on a releative tangent... You have a lot more experience with EC2 than I; do you (or others) have any theory as to why there is such a huge discrepency in pricing between us-east-1d (currently ~$0.1342) and us-east-1c (~$2.10)? That's over 15 times! I know there have been some studies on this; do you (or anyone) have any good references? (I'm just intriqued is all.) 2014-11-19, 19:39 #3163 Primeinator "Kyle" Feb 2005 Somewhere near M50..sshh! 2·3·149 Posts Quote:  Originally Posted by chalsall Yup. Just sign up to EC2, spin up an instance which can support NVideo CUDA (there are a bunch free ones; choose your faviorate distro) on a cg1.4xlarge instance type, install CUDA (and apply security updates, etc). Next, create a new 1GB volume, and amoung other customizations change /etc/fstab to include a line which reads "/dev/sdf1 /home ext4 noauto,noatime 1 0", and then attach this to your instance (by default it will be "/dev/sdf"). Use "fdisk /dev/sdf" to create a single linux partition, and then "mkfs.ext4 /dev/sdf1". Now, copy (with the archive flag) the existing "/home/ec2-user" directory to "/tmp/" Lastly "mount /home" should bring the newly created partition online. Move (or copy) the "/tmp/ec2-user" directory to "/home". "shutdown -r now", log back into the instance, and execute "sudo mount /home". Make sure the new home partition mounts correctly. If it does, then install all your desired mfaktc software in two seperate directories (including any submission / fetching spiders desired, and create the related cronjobs) and give it a go. Launch two mfaktc instances (the second with "-d 1"). You may also want to install an instance of mprime as well. If everything's working the way you want then "Create Image" from the "Instances" tab. This will create a custom AMI which you'll boot from in the future (thus avoiding the need to install CUDA and the software each time. Note that, importantly, the AMI file system is ***NOT*** preserved between instance creations; this is the reason for the extra volume created for "/home" -- mounted volume's are preserved. Oh, and know that spot instances can go away at any time without notice. But, they're dirt cheap; I usually bid at$0.15 per hour, and often have one (or more) instances up for many days paying on average $0.135 an hour. It's actually relatively easy to do. Thanks! I am going to look into this. 2014-11-19, 20:28 #3164 Mark Rose "/X\(‘-‘)/X\" Jan 2013 Ͳօɾօղէօ 3·72·19 Posts Quote:  Originally Posted by chalsall Mark: on a releative tangent... You have a lot more experience with EC2 than I; do you (or others) have any theory as to why there is such a huge discrepency in pricing between us-east-1d (currently ~$0.1342) and us-east-1c (~$2.10)? That's over 15 times! I know there have been some studies on this; do you (or anyone) have any good references? (I'm just intriqued is all.) Amazon has never said specifically the algorithm behind the spot pricing. People who have looked more into it than I have come up with some speculations that make sense though. Spot pricing is basically for unused reserved instances. People buy reserved instances for two reasons: one, if running most of the time, they save money; and two, they get first dibs on capacity, ahead of on-demand and spot. However, they don't have to be fully using their reserved instances. These unused instances can be sold on the spot market. Amazon sets a floor on spot capacity, my guess is to cover the electrical cost of having the machine run. If there are sufficient spot requests to run all the spot capacity for an instance type, the highest bids win. Some people may set a higher than on-demand spot price to take advantage of low pricing most of the time, while also have a better chance to keep their instance running if capacity becomes constrained. If all the reserved instances are in use, and the least-winning spot bid is at the on-demand price, on-demand capacity can be used to fulfill spot requests. The on-demand price for a cg1.4xlarge is currently$2.10 in us-east-1. In the one zone that has no spot capacity (probably no instances are reserved, because most people are using the g2.2xlarge instances), so the spot price is the on-demand price.

In the other two zones, one has a slightly higher bid price, likely because someone is willing to pay more to keep their spot instances in the same zone (with other spot instances or otherwise).

In eu-west-1 my guess is no one has unused reserved instances, so there is no spot capacity.

 2014-11-19, 23:13 #3165 TheMawn     May 2013 East. Always East. 6BF16 Posts So what exactly happened to that one-year buffer we had? I never really understood how we suddenly had a larger demand for DC's...
2014-11-19, 23:29   #3166
Mark Rose

"/X\(‘-‘)/X\"
Jan 2013
Ͳօɾօղէօ

3×72×19 Posts

Quote:
 Originally Posted by TheMawn So what exactly happened to that one-year buffer we had? I never really understood how we suddenly had a larger demand for DC's...

George changed it so that new users get DCLL instead of LL by default for various reasons. Instead of 50ish DC being requested per day, it jumped to 1000 per day. 90% of those will be recycled, so the rate will drop to 150ish per day, after 60 days (mid January).

2014-11-20, 02:30   #3167
Uncwilly
6809 > 6502

"""""""""""""""""""
Aug 2003
101×103 Posts

26×3×43 Posts

Quote:
 Originally Posted by chalsall Yup. Just sign up to EC2, spin up an instance which can support NVideo CUDA (there are a bunch free ones; choose your faviorate distro) on a cg1.4xlarge instance type, install CUDA (and apply security updates, etc).
If I could understand this (all), I would likely spend about $30 a month toward the 100M digit range. 2014-11-20, 13:37 #3168 chalsall If I May "Chris Halsall" Sep 2002 Barbados 235016 Posts Quote:  Originally Posted by Uncwilly If I could understand this (all), I would likely spend about$30 a month toward the 100M digit range.
It really is quite easy. I'd suggest you sign up for EC2, and experiment with their "Free Teir" offer; gives you two "t1.micro" instances (one Linux and one Winblows) to experiement with for a year for free. Not powerful enough to do much LL'ing (and no GPU'ing), but a good way to get comfortable with the UI, creating and mounting volumes, etc.

If you're interested, I could probably figure out how to export my AMI (already configured with CUDA) for you to use.

 Similar Threads Thread Thread Starter Forum Replies Last Post Primeinator Operation Billion Digits 5 2011-12-06 02:35 1997rj7 Lone Mersenne Hunters 27 2008-09-29 13:52 Uncwilly Operation Billion Digits 22 2005-10-25 14:05 paulunderwood 3*2^n-1 Search 2 2005-03-13 17:03 1997rj7 Lone Mersenne Hunters 25 2004-06-18 16:46

All times are UTC. The time now is 17:37.

Thu Jul 16 17:37:42 UTC 2020 up 113 days, 15:10, 2 users, load averages: 1.78, 1.50, 1.60