General Concepts & Problems
Q: What is High-Performance Computing (HPC)?
A: HPC is using parallel processing to run advanced applications that are either too large or would take too long for standard computers. It's a network of nodes, each with processing units and its own memory. As computer performance has increased, power consumption has also significantly increased. This leads to several problems.
Q: What are the three main categories of problems?
Economical Impact:
High electricity bills (e.g., US desktop computers >10% of electricity).
Large data centers (e.g., Google) consume as much electricity as a small town.
Cooling can account for up to 70% of energy costs.
Supercomputers might need entirely new buildings.
Ecological Impact:
Significant carbon dioxide (CO2) production (desktops like millions of cars).
Worsening heat dissipation problems.
IT industry's greenhouse gas emissions comparable to international airlines.
Technical Impact:
Large multiprocessors need expensive packaging and cooling.
Increased heat leads to a higher demand for fault-tolerant systems.
Power Reduction Strategies & Techniques
MCQ: Besides FLOPS (Floating Point Operations Per Second) for performance, the Green500 list uses FLOPS/Watt to rank the energy efficiency of computing systems.
Q: What are the two main approaches to reducing power consumption?
Thermal-aware Hardware Design: Reducing power consumption using logical techniques. Focuses on physical components (e.g., device-level power reduction, low-power processors, low-power memories).
Power-aware Software Design: Focuses on software control (e.g., operating systems, compilers, application-level power management).
Q: What is Dynamic Power Management (DPM) Dynamic Voltage and Frequency Scaling (DVFS)?
A: DPM uses software to control voltage regulation. The main idea is to monitor CPU idle time and reduce energy consumption by lowering clock speed and minimizing idle time. DVFS is the most commonly used technique for CPU Power management. It reduces power dissipation by lowering the processor's clock speed and supply voltage.
Power Allocation & Scheduling
Q: What are the four non-trivial sub-problems in power allocation and task scheduling?
System Partitioning: Deciding how to divide a multiprocessor into clusters for tasks.
Precedence Constraining: Handling dependencies between tasks, which makes designing algorithms harder.
Task Scheduling: Arranging how and when precedence-constrained tasks run.
Power Supplying: Applying appropriate power levels and execution speeds to tasks.
Q: What are the three main types of power-aware task scheduling algorithms?
Pre-Power Determination Algorithm: First determines power supplies, then schedules tasks.
Post-Power Determination Algorithm: First schedules tasks, then determines power supplies.
Hybrid Algorithms: Interleaves task scheduling and power supply determination.
Q: Can you give examples of Pre-Power Determination algorithms?
Equal Time Algorithm: Sets power so all tasks have identical execution time.
Equal Energy Algorithm: Sets power so all tasks consume the same amount of energy.
Equal Speed Algorithm: Sets power so all tasks execute at the same speed.
Q: What are the different levels of CPU Power Management discussed?
Q: What is Component-Based Power Management?
A: It acknowledges that while the CPU often uses the most power, other components (memory, disk, network) also consume significant power. This approach considers the power needs of each component, which can vary with the workload.
Q: How is power management addressed in Virtualized Datacenters?
A: With the growth of cloud computing, VMM (Virtual Machine Monitor) is used to control the coordination among Virtual Machines (VMs) to reduce power consumption in virtual datacenters.
Q: What is High-Performance Computing (HPC)?
A: HPC is using parallel processing to run advanced applications that are either too large or would take too long for standard computers. It's a network of nodes, each with processing units and its own memory. As computer performance has increased, power consumption has also significantly increased. This leads to several problems.
Q: What are the three main categories of problems?
Economical Impact:
High electricity bills (e.g., US desktop computers >10% of electricity).
Large data centers (e.g., Google) consume as much electricity as a small town.
Cooling can account for up to 70% of energy costs.
Supercomputers might need entirely new buildings.
Ecological Impact:
Significant carbon dioxide (CO2) production (desktops like millions of cars).
Worsening heat dissipation problems.
IT industry's greenhouse gas emissions comparable to international airlines.
Technical Impact:
Large multiprocessors need expensive packaging and cooling.
Increased heat leads to a higher demand for fault-tolerant systems.
Power Reduction Strategies & Techniques
MCQ: Besides FLOPS (Floating Point Operations Per Second) for performance, the Green500 list uses FLOPS/Watt to rank the energy efficiency of computing systems.
Q: What are the two main approaches to reducing power consumption?
Thermal-aware Hardware Design: Reducing power consumption using logical techniques. Focuses on physical components (e.g., device-level power reduction, low-power processors, low-power memories).
Power-aware Software Design: Focuses on software control (e.g., operating systems, compilers, application-level power management).
Q: What is Dynamic Power Management (DPM) Dynamic Voltage and Frequency Scaling (DVFS)?
A: DPM uses software to control voltage regulation. The main idea is to monitor CPU idle time and reduce energy consumption by lowering clock speed and minimizing idle time. DVFS is the most commonly used technique for CPU Power management. It reduces power dissipation by lowering the processor's clock speed and supply voltage.
Power Allocation & Scheduling
Q: What are the four non-trivial sub-problems in power allocation and task scheduling?
System Partitioning: Deciding how to divide a multiprocessor into clusters for tasks.
Precedence Constraining: Handling dependencies between tasks, which makes designing algorithms harder.
Task Scheduling: Arranging how and when precedence-constrained tasks run.
Power Supplying: Applying appropriate power levels and execution speeds to tasks.
Q: What are the three main types of power-aware task scheduling algorithms?
Pre-Power Determination Algorithm: First determines power supplies, then schedules tasks.
Post-Power Determination Algorithm: First schedules tasks, then determines power supplies.
Hybrid Algorithms: Interleaves task scheduling and power supply determination.
Q: Can you give examples of Pre-Power Determination algorithms?
Equal Time Algorithm: Sets power so all tasks have identical execution time.
Equal Energy Algorithm: Sets power so all tasks consume the same amount of energy.
Equal Speed Algorithm: Sets power so all tasks execute at the same speed.
Q: What are the different levels of CPU Power Management discussed?
OS-Level Power Management
- It puts the CPU to sleep when it's not needed (C-States).
-
It adjusts the speed and performance of the CPU depending on what you're doing (P-States).
-
Modern systems use something called ACPI to manage this automatically
Workload-Level Power Management
- It uses smart scheduling and planning to save power.
-
For example, if you're just typing, it uses less CPU. If you're gaming, it gives more power.
Cluster-Level Power Management
- A special tool watches all servers and makes sure they don’t use too much power.
-
If any server uses too much, it slows them down to stay within the agreed power limits (SLA = Service Level Agreement)
Q: What is Component-Based Power Management?
A: It acknowledges that while the CPU often uses the most power, other components (memory, disk, network) also consume significant power. This approach considers the power needs of each component, which can vary with the workload.
Q: How is power management addressed in Virtualized Datacenters?
A: With the growth of cloud computing, VMM (Virtual Machine Monitor) is used to control the coordination among Virtual Machines (VMs) to reduce power consumption in virtual datacenters.
Comments
Post a Comment