Summer 2013 and winter 2014, there were an inordinate number of power outages in Baker Lab, and other Chem buildings! |
ChemIT's record of recent power outages
Date | Outage duration | Cause | Official link | ChemIT notes |
|---|
2/27/2014
Thursday | 2 minutes
CU's record: Per 3/3/14 email: On Thursday 2/27/14 at 9:50am the campus experienced a disruption of electricity. The whole campus experienced an approximately 30 second outage. Many buildings on the central, west and north campus areas remained without power for 30 minutes to an hour.
Oliver's record: Power outage at 9:53a. Restored at 9:55a (more than one minute). | Procedural error?
Per 3/3/14 email: The outage was caused as a result of routine maintenance activities which were being conducted at the Campus’ main substation which takes power from the NYSEG transmission system and provides it to campus. This work has been conducted many times before without incident but in this case caused a major disruption of electricity supply. Staff from Utilities and external technical resources are investigating the root cause of this unexpected event. | No link to info in 3/3/14 email?
Some delayed time-stamps, from the IT folks:
http://www.it.cornell.edu/services/alert.cfm?id=3072
Initial timing info, from the power folks:
http://www.cornell.edu/cuinfo/specialconditions/#2050 | Michael led our effort to initially evaluate and restore systems, with Oliver adding to documentation and to-do's. Lulu completed the cluster restoration efforts.
Lost 1-2 hours each for Michael, Lulu, and Oliver.
Cornell called it a "power blip". In Oliver's books, any outage longer than seconds is not a "blip".
Q: Broke one GPU workstation? |
1/27/2014
Monday | 17-19 minutes
CU's record: Power outage at 2:22p. Restored around 2:41p.
Oliver's record: Power outage at 2:22p. Restored at 2:39p. | ? | http://www.it.cornell.edu/services/alert.cfm?id=3040 | Lulu, Michael, and Oliver shut down head nodes and other systems which were on UPS. (Those systems non UPS shut down hard, per usual.)
Lost 3 hours, for Lulu, Michael, and Oliver.
Roger away on vacation (out of the U.S.) |
12/23/2013
Monday | 2 minutes
CU's report: 08:36 AM - 8:38 AM
(ChemIT staff not in yet.) | Procedural error? | http://www.it.cornell.edu/services/alert.cfm?id=2982 | Terrible timing, right before the longest staff holiday of the year.
ChemIT staff not present during failure.
Lost most of the day, for Roger and Oliver.
Michael Hint and Lulu way on vacation (out of the U.S.) |
7/17/13 | Half a morning (~2 hours)
CU's report: 8:45 AM - 10:45 AM | | http://www.it.cornell.edu/services/alert.cfm?id=2711 | |
Question: When power is initially restored, do you trust it? Or might it simply kick back off in some circumstances?
- Because we don't know the answer to this question following any specific power outage, we are reluctant to turn back on servers right away. Instead, we like to wait ~10-25 minutes.
What would it cost to UPS our research systems?
Assuming protection for 1-3 minutes MAXIMUM:
Do all head nodes and stand-alone computers in 248 Baker Lab
- Started getting done. About $180 (APC brand) per head node or server every ~3-4 years (3 yr for warranty and ~4 years actual battery life). And ~$900/ set of 4 GPU systems every ~3-4 years (to confirm approach and estimates).
CCB head nodes' UPS status:
Remaining UPS's to invest in
Clusters
Most we been done Spring 14, after the spate of power failures. See CCB's HPC page (first chart, in "UPS for headnode" column) for details
Cluster | Done | Not done | Notes |
|---|
Loring | | X | Unique: Need to do ASAP |
Abruna | | X | Unique: Need to do ASAP |
Non-clusters
See CCB's HPC page (second chart, in "UPS" column) and CCB's non-HPC page (in "UPS" column) for details of the few that are already done.
Stand-alone computers' UPS status:
Computer | Done | Note done | Notes |
|---|
Coates: MS SQL Server | | X | Unique: Need to do ASAP |
Freed: Eldor | | X | Unique: Need to do ASAP? (Q: Is OS backed up?) |
Baird: 1 rack-mounted computational computer | | X | Need? |
| | | |
Review others at above two cited pages which might need a UPS, after above ones done.
Switches
Do all switches: Maybe ~$340 ($170*2), every ~4 years.
- Recommend: Do ASAP.
- Q: Funding?
- Other issues and concerns with actually implementing this approach:
- Rack space. Maybe power. Maybe cord lengths. What other issues?
Do all compute nodes: ~$18K initially, and perhaps ~$4.5K every ~4 years to replace batteries and deal with UPS hardware failures.
- ~20 20amp UPS's ($900 each) required.
- Replacement batteries ~$200 each, or ~1/4 replacement cost.
- Estimates are simply back-of-the-envelope calculations.
- If were to actually implement, there may be smarter ways to do this, but the total cost will likely not be lower.
- In fact, costs may be higher, if sufficiently higher benefit doing it a different way, for example.
- Issues and concerns with actually implementing this approach:
- Costs. Rack space. Maybe power. Maybe cord lengths. What other issues?
Compute node counts, for UPS pricing estimates. Does not include head node:
Cluster | Compute node count | Power strip equivalents
(~8/strip MAX) | Cost estimate,
every 4 years | Notes |
|---|
Collum | 8 | 1 | $900 | |
Lancaster, with Crane (new) | 10 | 2 | $1.8K | |
Hoffmann | 19 | 2 | $1.8K | |
Scheraga | 91 | 13 | $11.7K | |
Loring | 4 | 1 | $900 | |
Abruna | 9 | 1 | $900 | |
C4 head node: pilot
Widom's 2 nodes there. | N/A | N/A | N/A | This CCB Community head node pilot has no compute nodes of its own.
It hosts compute nodes from CCB researchers. |
Widom | 2 | 1 | ? | Compute nodes are hanging off of "C4" head node, above. |
TOTALS | ~140? | 21 | $18K + Widom | |
Procedures and reminders
Reboot switches?
- 8 switches support the clusters (out of 12 in the room).
Start head nodes, if not on already
- Only a few are on UPS. Those can obviously be left on.
- None should be set to auto-start on power-off.
Confirm head nodes accessible via SSH
PuTTY on Windows
Use FMPro to get connection info?! (not the right info there, though...)
Menu => List Machine / User Selections = > List Group or selected criteria
- Machine function => Server HN
Start compute nodes
If nodes done show up, consider:
- Restart Torque scheduler on problematic nodes.
- Try rebooting the switch the affected nodes are connected to, especially if the problematic nodes are grouped to a single switch.
- Hook up a monitor as one of the high nodes boot.