This paper proposes a new approach, Fairness via Source Throttling (FST), to provide fairness in the entire shared memory system of a chip-multiprocessor (CMP) system by eliminating the need for developing fairness mechanisms for each individual resource. FST ensures fairness decisions are made in tandem in the entire memory system, enforces thread priorities/weights, and allows system software to enforce different fairness objectives and fairness-performance tradeoffs in the memory system. The evaluations show that FST provides the best system fairness and performance compared to four systems with no fairness control and with state-of-the-art fairness mechanisms.
Strengths
- I really appreciate the figure in the introduction section, which tells a very impactful story with an simple case of how quality of service is influenced by the unfairness caused by resouce sharing.
- The idea of uniformed resources (namely source in this paper) is the key contribution of this work, which raised the fairness controlling from the level of various resources to the upper level of CPU cores. It noted and solved the conflicts of differerent fairness mechanisms in precious works.
- The idea of estimating the interference is crutial to the mechanism
Weaknesses
- Security problem: the dynamic machenism poses a thread on the security side of the fairness controller when the attacker could issue specifically designed requests to resource on a core to slow down the tasks on the other cores.
- The idea of measureing/estimating unfairness in previous interval and adjusting/throttling in the next interval is cool, but how to design the measurement interval is also a challenge to this mechanism. This mechanism will cause delay between intervals, and small tasks in big intervals may cause waste in time for computing the unfairness rate.
Can you do better?
- I would do some measurement over the conflicts between os level optimization of the FST mechanism.
- Why the author choose to separatly tell the results on 2-cores and 4-cores. It would be better to combine these two sets of results as there is no essential difference between these two experiment settings.
- Continuing the experiment setting problem, I would like to add more setting with different number of cores, such as 8, 16. Multicore is really common on computers (even on PC, leaving alone huge servers). Why not demonstrate the effect of fairness control on many-core machines?
Takeaways
- Sometimes estimating works just as fine as precise calculation, which usually seems not practical. Always remember that our goal is to improve the performance rather than measuring the performance.
- As mentioned in the strengths, learning to raise some thing from one level from the system to another upper level gives us a better view and possibly greater performance.
- It is the first paper among all the serials I have seen that includes case study. Case study help readers understand special cases and poses explanations over some insteresting results.
Other Comments
This research presents an efficient architectural method for ensuring fairness in CMP systems by enabling fair sharing of the entire memory subsystem. Unlike prior methods, this technique does not require multiple complex and specialized fairness mechanisms for different shared resources, making it more cost-effective and easier to implement.