Paper Title:- Adrenaline: Pinpointing and Reining in Tail Queries with Quick Voltage Boosting

Problems to solve

  • Reduce tail latency (95ile) cut off times.
  • Given a tail latency target,improve energy efficiency
  • Current DVFS techniques not sufficient to get the job done

Difference with PEGASUS

  • This paper aims at reducing tail latency of leaves where as PEGASUS aimed at latency at the root node.
  • Motivation of both the papers are same, However the problem solved is different.PEGASUS used available latency to improve energy whereas this paper aims to reduce tail latency.

Users

  • End users of internet. They get better quality
  • Operators of datacenters.

Solution

  • Identify queries which can be boosted as the packet is received.
  • Use short step to switch between low and high voltage to boost or not boost the query.
  • A coarse gain DVFS was also employed to change the voltage of the low and high rail themselves. It was based on the workload but the workload was not clearly defined in the paper.

Unique about the solution

  • Boosting specific tasks which can benefit from boosting.
  • Predicting which tasks to boost.
  • Rely on new technology which enables fine grain boosting.
  • Coarse grain DVFS for load
  • Clock distribution (This part was not clearly explained in the paper)

Technology’s role

Short step benefits from technological changes as the size of capacitors can be reduced to achieve low area and fast switching. This was used to obtain per query switching.

Power modelling

Following equations based on heuristics were used to model the power

P0 = P0dyn + P0static
Pdyn ~ proportional to V2f ~ f3

In paper Pdyn/P0dyn = (f/f0)2.7

Pstatic ~ V ~ f

In paper Pstatic/P0static = (f/f0)

P0static ~ = 0.2*P0static (used for circuits validity for processors not certain as caches consume same power always.)

Evaluation

Analysis
* Web search and memcached workloads were run at different frequencies to identify most common request types which contribute more to tail latency.
* Analysis also helped to identify which queries benefit from boosting. SET for memcached workload and short length queries for web search.
  • Simulated a cluster in Big house simulator.
* big house uses service time distribution length of a task and arrival time of the task to simulate a cluster.
  • Analytical Power model.
  • queuing delay was not added. Latency
  • Although aim of the paper was to reduce tail latency at end user, it was not clearly mentioned what latency has been included.

Typical process is User | Internet | Sending to leaves & Queuing | Process | Collect | Internet | Users. Which part of the latency is being aimed by the paper was not clearly stated.

For a baseline, DVFS, and Adrenaline following were measured

* Measured 96ile latency, energy, distribution. All these were compared
* Pareto optimal curve
* Given a target latency and slack. measured saving in energy. Started at max energy here.
  • Measurements taken were such that they maximize the benefit of the Adrenaline e.g. starting at low energy for baseline, DVFS.

Verdict

Although paper had some unique solution like identifying which tasks to boost but it failed to establish trust with readers.