Posts by Garry

1) Message boards : Number crunching : Posts moved from the FAQ thread (house cleaning) (Message 11637) Posted 4 Mar 2006 by Garry Post: Why should I set BOINC to keep application in Memory? If the application is removed from memory during an application swap, it will loose the work performed since the last checkpoint. In the case of Rosetta the checkpoints occur each time the percentage advances. Since is takes nominally 90-120 min between 10% checkpoints, if you do not keep the application in memory, and you set the swap interval to less than the time it takes your machine to reach a 10% mark, the work units can appear to be 'hung". This is why the recommended configuration is to set "Keep applications in memory" to "YES". As an added protection, it is also recommended to set a value of nominally 120 min between application swaps. Either of these settings alone has been shown to reduce the problem, but they can be used togeather for better results It should be noted that ALL BOINC projects suffer from loss of CPU cycles if the applications are not kept in memory. Any work that is not save at a checkpoint before a swap, is lost when a swap occurs. On applications like Rosetta and Climate prediction this loss is significant (15 min to over an hour). On projects like Predictor, SETI and Einstein, the loss is less but it is still there (usually 60 seconds using the default setting for writing to disk). People who wish to squeeze every cycle out of their machine, usually keep applications in memory for this reason. It is important to note that the memory we are talking about is not actual RAM but in fact is virtual memory (on disk), so the actual impact of this is not that significant. In addition you can adjust how much memory is used by the application in this regard. Question submitted by: Dimitris Hatzopoulos Team: Greece United Dimitris, My compliments on two pieces you've written in this FAQ, the one above and "What is project 'Swapping' and how should I have it set?". I'm learning from them. I value that. Additionally, I'm about to change settings on my BOINC, to include adding Rosetta back into the mix on my machine. (Don't get real excited about that. In today's world, this is a bit of a slow machine.) As a means of suggesting additional material to cover, possibly adding still more strength to your writeups, I'll offer these questions: - When BOINC decides on an application swap, putting Rosetta data on disk in virtual memory and the machine reboots before Rosetta gets the processor again, what happens to the data in virtual memory? Lost? Checkpoint just before BOINC closes out? - Is there some reason the Rosetta application doesn't do a checkpoint immediatly before an application swap? Is BOINC designed to order the application to do a checkpoint and delay the swap until the checkpoint is complete? - How many checkpoints does Rosetta take each work unit? Your writeup says, "In the case of Rosetta the checkpoints occur each time the percentage advances. Since is takes nominally 90-120 min between 10% checkpoints". As I read the first part of that, it sounds like Rosetta does roughly 100 checkpoints. The second part sounds like Rosetta does roughly 10 checkpoints. - In the article, "What is project 'Swapping' and how should I have it set?", you include, "Once set, this preference will apply to all active projects on your machine. For this reason you may see some performance issues on your system." What performance issues? What range of performance change should I expect? What characteristics of my machine determine the particular performance change? So often, if one person has these questions, many more will benefit from seeing the same answer. I have them; I offer them to you hoping I'm helping. Thanks in advance for any answers you decide to include. Regards, . . . Garry
2) Questions and Answers : Wish list : more frequent checkpointing (Message 8172) Posted 2 Jan 2006 by Garry Post: Great project. However, I'm not comfortable with the need to keep data in memory. It seems like you could get past that by augmenting your checkpointing code, making it easier and more effective to save the current state of work. It seems like you'd want to run that checkpointing code just before the computer shuts down (if you're running then) and just before you relinquish the processor to another BOINC project. Maybe there's information that'll make me more comfortable with leaving data in memory. Do you tie up only a "small" amount of memory? Does Windows put that data in the swap file on disk if you don't happen to be running?
3) Questions and Answers : Windows : Aborted work unit and memory usage (Message 3773) Posted 20 Nov 2005 by Garry Post: I saw a recommendation in the system requirements that Rosetta should be allowed to stay in memory when not running. And that the code does a checkpoint four times or so during each work unit. I suspected that if I didn't let it stay in memory, it would lose all work since the last checkpoint. I tested. The first time BOINC gave the processor to another experiment (and kicked Rosetta out of memory), Rosetta reported "Result 1n0u__abrelaxmode_random_length20_jitter02_omega_sim_aneal_bab100_12350_0 exited with zero status but no 'finished' file". CPU time and progress reported zero. Is it reasonable to assess that Rosetta didn't do a checkpoint during the time it ran on my machine? And that the time my machine contributed to Rosetta was lost?