Affected
- ResolvedResolved
This issue has become one of supply and availability. As such, we are changing this issue to Maintenance as there is nothing further we can do until the vendor replaces the hardware. Please continue to use other GPU resources. We will notify the community once we have new hardware to replenish the gpu_test queue.
Please see Running Jobs for other available resources.
- IdentifiedUpdate
Difficulty obtaining replacement mainboards has this issue stalled. More information when we have it.
- IdentifiedUpdate
Correction to first part of previous update:
Currently only 2 of the nodes in gpu_test are working and the single remoteviz node is down. These will require help from the vendor to revive. We are also investigating with them the root cause for multiple node failures. No ETA
- IdentifiedUpdate
Currently only 2 of the nodes in gpu_test are working and the single remoteviz node is down. These will require help from the vendor to revive. We are also investigating with them the root cause for multiple node failures. No ETA
As a result of this, we are pre-emptively modifying the QoS for gpu_test now. We had planned to make a change on Thursday after announcing it in tomorrow's maintenance email, but this forces our hand.
QoS changes for gpu_test effective now: Limited to 1 job per user. That job is limited to a maximum 16 cores and 90GB memory.
See Running Jobs for a list of available partitions, including the gpu partition.
- IdentifiedIdentified
Please note that this issue also affects the remoteviz partition.
Updates as we have them.
- InvestigatingInvestigating
Several hosts in the gpu_test queue have become unresponsive and will require a physical visit to reset. Staff are en route to the data center.
On a related note, please do not use gputest as a workaround for the regular gpuqueues.
This is unfair to other users and ties up the gpu_test partition. This partition is not for general job use.We will be addressing this issue later this week by reducing the number of jobs allowed per user. This will be noted in tomorrow's maintenance email and then implemented Thursday.