Equipment configurations are always subject to safety as first consideration. If it cannot be done safely then it should not be done. Consult SCC Supervisor when in doubt.
Power Budget (3KW)
The rule of the competition is to power the equipment used in an HPC cluster within the power limit of 3000W (approximately ~13A) on 1 power circuit while running the benchmarks and applications specified. The power distribution unit (PDU) that will be used to monitor the power usage for each team cannot exceed that limit.
Each team will be provided with two power circuits in the booth. One socket from one circuit will be used for the APC PDU that used to power the equipment for the competition, the other circuit can be used for power other miscellaneous equipment that the team might have.
There will be two Types of power outlets (at least one of the outlets is rated at 16A) will be provided for each team. Only 1 PDU will be given to each team which will be plug into 1 of the power outlets. The other Type F power outlet will have a power strip. Each team will be responsible for bringing in accessories, such as power strip (that takes the Type F plug), power cables for electric equipment, computer monitors, splitter cables (with proper plugs), etc. If you need to connect more devices, the team will need to bring in additional power strip. Any additional power strips, power plug converter, and power extension cords, splitter cables (with proper plugs), or other equipment needs to be supplied by the teams.
Please note that computer monitor can be excluded from the power budget.
For each circuit, there will be circuit breaker will have a limit of 16A. The Type F (Schuko) plugs are available to plug into the circuits.
The power distribution unit (PDU) used to monitor the power usage for the competition supports drawing of a maximum of 16A of power. We are using the APC AP8858 PDU to monitor the power usage through its Ethernet interface via SNMP for each team. The TypeF wall plug from each booth can supports up to a maximum of 16A power. As the rule states, each team will use the power cap at 3kW. Teams might need to prepare an adapter to convert from the TypeF (Schuko) wall power to the C20 socket for the input of the PDU.
The technical specifications for the competition PDU is detailed here:
There are 18 outlets for C13 and 2 outlets for C19 available. If more devices are needed to be plugged into the PDU to account for the power budget, the team will need to prepare and bring enough splitter cables (IEC C14 to C13 Splitter power cables) to power the additional devices. If the type of the plugs used for the machines are not taking C13, the teams will be responsible for bringing suitable power cables.
We will collect power consumption from the PDU to monitor the power usage by the teams. As soon as the power consumption is over 3000W, the overload warning light on the PDU will be illuminated to notify the team has exceeded the power budget. The SCC supervisors will be notified too that the power exceeded the limit. Points will be deducted each time the power budget is exceeded.
If the power consumption get well beyond the 3000W limit, the circuit breaker may trip and you will lose power to the systems. The team will also lose valuable system time to run the dataset while the circuit breaker is being reset, so the team will need to control the power draw throughout the competition.
The AP8858 PDU will display the actual wattage on the LCD indicator of the PDU. A link to the web page that shows the power draw by each team will be provided during the competition. As soon as the power draw exceeds 3000W by a team during the actual competition, the team will trigger the warning on the PDU to notify the team has exceeded the power budget. The SCC supervisors will be notified that the team has exceeded the power limit and ask the offending team to stop and rerun any application/benchmark runs might have run during the power limit. SCC supervisors may disqualify one or more run submissions if they observe that the power consumption is beyond the limit for an extended period of time. The idea is for teams NOT to exceed the 3000W limit. The team needs to control the power draw to 3000W or less.
Examples of Equipment In Power Budget
A laptop/desktop machine that does system monitoring is ok not to include in the power budget. The power budget covers all of the equipment that is actively used for the applications and benchmarking (for example servers. Monitors for example is not part of the budget.
However, if you plan on using a system provisioning software with management, that is responsible for distributing the software images/packages to the compute nodes. That will count as a part of the cluster and should count as a part of the power budget.
The rule is, if it has any direct impact on the application performance, and we need to account that for the power budget.
Using external/stored power source is not allowed. Such as using laptop battery or battery backup (for instance, Uninterruptible Power Supply or UPS) is not allowed.
Examples of power usage that counts toward power budget:
– Network switches
– System provision nodes
– Laptop/desktop running system monitoring software for provisioning system images.
– File system server (such as NFS server)
– Fans, and other method of cooling (such as liquid cooling)
Examples of power usage that does not count towards power budget:
– Laptop/desktop running system monitoring software for the power usage.
– Personal laptops or other equipment
Hands off System Throughout the Competition
No one can touch the equipment physically after the competition starts. If there is a need to touch the equipment an official SCC supervisor needs to be called and will rule on the issue. The only exception is in safety – if an unsafe condition is found, anyone can power down the equipment, and SCC supervisor must be called immediately afterwards.
No changes in the BIOS (or in EFI shell) are allowed after the competition has started. The systems are to be powered on at all times and no rebooting. System reboots can only reboot when the system is hung or it has encountered major failures.
Powered On at All Times and No Rebooting
All equipment used for the competition must be turned-on at all time (i.e. one cannot run one benchmark on part of the system and then power up the entire system for other benchmarks). System reboots are only necessary for hung or failed hardware. SCC Supervisor must be notified before rebooting any hardware.
Modes like hibernation or suspension (either suspend-to-RAM or to disk) are not allowed. Those modes do not qualify for the rule “power on at all times and no rebooting” because power on mode implies that compute nodes need to be in the operating mode throughout the competition. Hibernation, suspension, and sleep modes are not in the operating mode, they are in the Standby mode, thus not qualify according to the rule.
Using the cluster after submitting the daily results
The teams that submitted the daily results, allows to continue and use their cluster and even cross power.
The teams are not allowed to run jobs after the end of the time deadline, the servers should remain idle.
Competition Input Files
The input files along with the secret application and any other special instructions on running the applications dataset will be given at the beginning of the day to each team using USB before the competition starts on each day. For some applications, even the input files may be small, there are steps involved, such as domain decomposition, which will create additional data files for each CPU to process. Thus generate much more space than just the input files that are given. Teams will need to be familiar with the benchmark applications to understand the typical data set and the runtime IO requirements in order to come up with a good enough disk storage solution.
Teams may try to run benchmark dataset and benchmark suites that for their practice purposes. As for preparation, we expect teams to build the applications and run a few benchmarks to get familiar with the application and dataset to discover the characteristics of an application and datasets to their own advantages.