Both sides previous revision
Previous revision
Next revision
|
Previous revision
|
cluster:189 [2020/01/13 13:00] hmeij07 [General] |
cluster:189 [2024/02/12 16:47] hmeij07 [Priority Policy] |
==== History ==== | ==== History ==== |
| |
In 2006, 4 Wesleyan faculty members approached ITS with a proposal to centrally manage a high performance computing center (HPCC) seeding the effort with an NSF grant (about $190K, two racks full of Dell PE1950, a total of 256 physical cpu cores on Infiniband). ITS offered 0.5 FTE for a dedicated "hpcadmin". An Advisory Group was formed by these faculty plus hpcadmin (5 members, not necessarily our current "power users"). Another NSF grant reward was added in 2010 (about $105K). An alumni donation followed in 2016 (about $10K). In 2018 the first instance of "faculty startup monies" was contributed to the HPCC (about $92K, see "Priority Policy" below). In 2019, a TrueNAS/ZFS appliance was purchased (about $40K, [[cluster:186|Home Dir Server]]) followed in 2020 by a GPU expansion project (about $96K, [[cluster:181|2019 GPU Models]]). The latter two were self-funded expenditures, see "Funding Policy" below. To view the NSF grants visit [[cluster:169|Acknowledgement]] | In 2006, 4 Wesleyan faculty members approached ITS with a proposal to centrally manage a high performance computing center (HPCC) seeding the effort with an NSF grant (about $190K, two racks full of Dell PE1950, a total of 256 physical cpu cores on Infiniband). ITS offered 0.5 FTE for a dedicated "hpcadmin". An Advisory Group was formed by these faculty plus hpcadmin (5 members, not necessarily our current "power users"). Another NSF grant reward was added in 2010 (about $105K). An alumni donation followed in 2016 (about $10K). In 2018 the first instance of "faculty startup monies" was contributed to the HPCC (about $97.4K, see "Priority Policy" below). In 2019, a TrueNAS/ZFS appliance was purchased (about $40K, [[cluster:186|Home Dir Server]]) followed in 2020 by a GPU expansion project (about $96K, [[cluster:181|2019 GPU Models]]). The latter two were self-funded expenditures, see "Funding Policy" below. To view the NSF grants visit [[cluster:169|Acknowledgement]] |
| |
The Advisory Group meets with the user base yearly during the reading week of the Spring semester (early May) before everybody scatters for the summer. At this meeting, the hpcadmin reviews the past year, previews the coming year, and the user base are contributing feedback on progress and problems. | The Advisory Group meets with the user base yearly during the reading week of the Spring semester (early May) before everybody scatters for the summer. At this meeting, the hpcadmin reviews the past year, previews the coming year, and the user base are contributing feedback on progress and problems. |
Several months later a pattern emerged. The Provost would annually contribute $25K **//if//** the HPCC user base raised $15K annually in contributions. These funds would "roll over". That would amount to $160K in 4 years, enough for a hardware refresh or new hardware acquisition. Finances also contributed $10K annually for maintenance such as failed disks, network switches, etc ... but these funds do not "roll over". Use it or loose it. All fund cycles restart July 1st. | Several months later a pattern emerged. The Provost would annually contribute $25K **//if//** the HPCC user base raised $15K annually in contributions. These funds would "roll over". That would amount to $160K in 4 years, enough for a hardware refresh or new hardware acquisition. Finances also contributed $10K annually for maintenance such as failed disks, network switches, etc ... but these funds do not "roll over". Use it or loose it. All fund cycles restart July 1st. |
| |
In order for the HPCC user base to raise $15K annually, CPU and GPU hourly usage monitoring was deployed (using Openlava ''lsacct'' accounting). A dictionary is maintained listing PIs with their associated members (student majors, lab students, grads, phd candidates, collaborators, etc). Each PI then quarterly contributes to the user fund based on a scheme supposedly yielding $15K annually. | In order for the HPCC user base to raise $15K annually, CPU and GPU hourly usage monitoring was deployed (using scripts parsing the ''lsb.acct'' file). A dictionary is maintained listing PIs with their associated members (student majors, lab students, grads, phd candidates, collaborators, etc). Each PI then quarterly contributes to the user fund based on a scheme supposedly yielding $15K annually. |
| |
Here is queue usage for 2019 [[cluster:188|2019 Queue Usage]] and below is listed the 2019 contribution scheme. | Here is queue usage for 2019 [[cluster:188|2019 Queue Usage]] and below is listed the 2019 contribution scheme. |
A gpu hour of usage is 3x the cpu hourly rate.\\ | A gpu hour of usage is 3x the cpu hourly rate.\\ |
| |
We currently have about 1,450 physical cpu cores (all Xeon), 72 gpus (20x K20, 4x GTX2018Ti, 48x RTX2080S), 520 gb of gpu memory and 8,560 gb of cpu memory. Provided by about 120 compute nodes and login nodes. Scratch spaces are provided local to compute nodes (2-5 tb) or over the network via NFS (55 tb), consult [[cluster:142|Scratch Spaces]]. Home directories are under quota (10 tb) but these will disappear in the future with the TrueNAS/ZFS appliance (190 tb, 475 tb effective assuming a compression rate of 2.5x, consult [[cluster:186|Home Dir Server]]). A HPCC guide can be found here [[cluster:126|Brief Guide to HPCC]] and the (endless!) software list is located here [[cluster:73|Software Page]]. We run CentOS 6.10 or 7.6 flavors of OS. | We currently have about 1,450 physical cpu cores (all Xeon), 72 gpus (20x K20, 4x GTX2018Ti, 48x RTX2080S), 520 gb of gpu memory and 8,560 gb of cpu memory. Provided by about 120 compute nodes and login nodes. Scratch spaces are provided local to compute nodes (2-5 tb) or over the network via NFS (55 tb), consult [[cluster:142|Scratch Spaces]]. Home directories are under quota (10 tb) but these will disappear in the future with the TrueNAS/ZFS appliance (190 tb, 475 tb effective assuming a compression rate of 2.5x, consult [[cluster:186|Home Dir Server]] deploy in 2020). A HPCC guide can be found here [[cluster:126|Brief Guide to HPCC]] and the (endless!) software list is located here [[cluster:73|Software Page]]. We run CentOS 6.10 or 7.[6|7] flavors of OS. |
| |
| |
==== Priority Policy ==== | ==== Priority Policy ==== |
| |
This policy was put in place about 3 years ago to deal with the issues surrounding new monies infusions from for example; new faculty "startup monies", new grant monies (NSF, NIH, DoD, others), or donations made to the HPCC for a specific purpose (as in GTX gpus for Amber). All users have the same priority. All queues have the same priority (except the "test" queue which has the highest priority). Scheduler policy is FIFO overlayed by Round Robin. There is no "wall time" on any queue. | This policy was put in place about 3 years ago (2017) to deal with the issues surrounding new monies infusions from for example; new faculty "startup monies", new grant monies (NSF, NIH, DoD, others), or donations made to the HPCC for a specific purpose (as in GTX gpus for Amber). All users have the same priority. All queues have the same priority (except the "test" queue which has the highest priority). Scheduler policy is FIFO. There is no "wall time" on any queue. |
| |
There are few Principles in this Priority Access Policy | There are few Principles in this Priority Access Policy |
The second principle grants priority access to certain resource(s) for a limited time to a limited group. The same PI/users relationship will be used as is used in the CPU/GPU Usage Contribution scheme. Priority access specifically means: If during the priority period the priority members' jobs go into pending mode for more than 24 hours the hpcadmin will clear compute nodes of running jobs and force those pending jobs to run. This by now is an automated process via cron that checks every 2 hours. Steps involved are; find priority members' jobs pending for more than 24 hours, find a node with no priority members jobs running in that queue, close target node, requeue all jobs on that node, force pending job(s) to run, wait 5 mins, reopen node. | The second principle grants priority access to certain resource(s) for a limited time to a limited group. The same PI/users relationship will be used as is used in the CPU/GPU Usage Contribution scheme. Priority access specifically means: If during the priority period the priority members' jobs go into pending mode for more than 24 hours the hpcadmin will clear compute nodes of running jobs and force those pending jobs to run. This by now is an automated process via cron that checks every 2 hours. Steps involved are; find priority members' jobs pending for more than 24 hours, find a node with no priority members jobs running in that queue, close target node, requeue all jobs on that node, force pending job(s) to run, wait 5 mins, reopen node. |
| |
All users should be aware this may happen so please checkpoint your jobs with a checkpoint interval of 24 hours. Please consult [[cluster:147|BLCR Checkpoint in OL3]] (serial jobs) and [[cluster:148|BLCR Checkpoint in OL3]] (parallel jobs). | All users should be aware this may happen so please checkpoint your jobs with a checkpoint interval of 24 hours. Please consult [[cluster:190|DMTCP]]. |
| |
==== General ==== | ==== General ==== |
There are 557 lines in ''/etc/passwd'' at this writing. Assume 25 are system accounts, 25 collaboration accounts (hpc01-hpc25, can VPN, for non Wesleyan faculty/PI, AD accounts), 100 temporary/recyclable class accounts (hpc100-hpc200, can not VPN, local accounts) which then leaves a lifetime user base of roughly 400 user accounts. Of those, which come and go, 2 to 2 dozen users may be logged in at any time. | There are 557 lines in ''/etc/passwd'' at this writing. Assume 25 are system accounts, 25 collaboration accounts (hpc01-hpc25, can VPN, for non Wesleyan faculty/PI, AD accounts), 100 temporary/recyclable class accounts (hpc100-hpc200, can not VPN, local accounts) which then leaves a lifetime user base of roughly 400 user accounts. Of those, which come and go, 2 to 2 dozen users may be logged in at any time. |
| |
Rstore is a platform for storing research static data. The hope is to move static data off the HPCC and mount it read-only back onto the HPCC login nodes. 440 tb, fully replicated, is provided for this purpose (Supermicro storage boxes using Rsync as replication engine) for HPCC users and other Wesleyan groups. | Rstore is a platform for storing research static data. The hope is to move static data off the HPCC and mount it read-only back onto the HPCC login nodes. 440 tb, fully replicated, is provided for this purpose (Supermicro storage boxes using Rsync as replication engine). For HPCC users and other Wesleyan groups. |
| |
The Data Center has recently been renovated so the HPCC has no more cooling power problems (It used to be in the event of a cooling tower failure, within 3 hours the HPCC would push temps above 85F). No more. We have sufficient rack space (5) and power for expansion. For details on that "live renovation" process visit [[cluster:178|Data Center Renovation]]. It turned out the HPCC was consuming 1/3rd of all electric and cooling capacities. Go HPCC. | The Data Center has recently been renovated so the HPCC has no more cooling problems (It used to be in the event of a cooling tower failure, within 3 hours the HPCC would push temps above 85F). No more. We have sufficient rack space (5) and power for expansion. For details on that "live renovation" process visit [[cluster:178|Data Center Renovation]]. It turned out the HPCC was consuming 1/3rd of all electric and cooling capacities. Go HPCC. |
| |
| |