11/24 Reviews - CuttingTheBill

[Rodrigo]

Please post your reviews here.

[James Tavares]

November 24, 2009

*CuttingTheBill*

Paper Title: Cutting the electric Bill for Internet-Scale Systems

Author(s): Asfandyar Qureshi, Rich Wber, Hari Balakrishnan, John Guttag,
Bruce Maggs

Date: SIGCOMM �09. August 17-21, 2009.

Novel Idea: The authors argue that cost savings in the millions of
dollars can be achieved by making data-center routing decisions based on
real-time energy costs, routing user requests to data centers which have
the lowest current cost per megawatt-hour.

Main Result(s)/Evidence: The authors use activity reports from Akami and
market pricing data from various regional ISOs to determine the cost
savings possible when using energy prices in the routing decision. The
authors find that energy elasticity has an enormous impact on
effectiveness; if idle servers consume the same amount of electricity as
loaded servers then it makes little sense to shuffle requests around.
Experiments also confirmed that cost savings could be achieved even when
bandwidth (at 95/5) and the maximum distance between client and server
are held as constraints. The authors do a good job of enumerating the
underlying assumptions to their model in their discussion.

Impact: Hard to tell; it seems intuitive to me that processing requests
in data centers with lower energy costs could save one money, at the
possible expense of bandwidth and latency.

Prior/Competitive Work: A formal discussion of prior work is notably
missing from the paper. The authors note in their introduction that
techniques which seek to reduce energy consumption (such as by
decreasing cooling costs, switching to DC power, using virtualization,
etc) are complimentary to their work.

Reproducibility: Any effort to reproduce the author�s results would
likely involve collecting utilization data from sources other than Akami
and looking for conclusions similar to those reached by the authors.
There is not much utility in reproducing the author�s experiments otherwise.

Question: The estimated energy costs cited for Google ($38M) account for
one-third of one-percent of its expenditures for 2007, or 0.22% of
revenues. Are energy costs really relevant?
[http://investor.google.com/fin_data.html]

Criticism:

1.) Assuming that a distributed system is �fully replicated� is a bit
na�ve� We�ve read many papers this semester where the opposite is true.
For example, a user profile is stored at one primary location and
replicated only to a few others.

2.) Does this work depend too heavily on the organization of the U.S.
energy market? The trend appears to be toward unifying the power grid �
will that promote national price parity and make parts of this work
irrelevant?

3.) I�m sure is incidental, but what is the energy penalty for running
these added algorithms?

4.) The findings in this paper benefit from hindsight (�if we did X, we
would have saved Y�). The paper acknowledges that most sites hedge power
costs by buying fixed-price contracts�these contracts are in effect
�insurance policies� against potentially unbounded price increases in a
volatile market, and in turn might have a value greater than the
estimated cost savings proposed by the authors work. The paper also
side-steps the question of whether companies can achieve better cost
savings in the long-term by negotiating fixed-price contracts vs using
the proposed real-time routing techniques.

Future Work:

1.) As data centers grow, could they turn to generating some or all
their own power? I believe this is a common practice in large factories
today.

2.) Figure out how I can sign up for some of those �negative energy prices�.

[Sunil Mallya]

Paper Title
Cutting the electric bill for internet-scale systems
Author(s)
Qureshi, Asfandyar, Weber, Rick, Balakrishnan, Hari, Guttag, John, and
Maggs, Bruce
Date
SIGCOMM 2009

Novel Idea
Proposing a method to reduce servicing costs in distributed systems by
taking advantage of the fact that different datacenters at any given
point of time have varying electricity cost.

Main Result(s)
The authors give us enough insight into electricity price variations
across different geographical locations over time and propose a new
method to reduce the energy costs in distributed internet scale
systems. Main factors that their method depends on are the fact that
electricity prices vary and large systems already have distributed
data geographically and each of these locations could handle requests.
Finally authors claim that this method could be easily integrated with
the existing frameworks to reduce costs.

Impact
Using Cost-Aware (Energy price conscious) routing, enterprises could
reduce costs could be reduced significantly in data centers.

Evidence
They analyze the electricity markets and generation resources, and
then they simplify the model and along with bunch of assumptions they
try to convince us that their model can be incorporated into the
datacenters and that would result in cost reductions.

Question
I think their model looks to be optimized for AKAMAI like heavily
geographically distributed systems, Simpler models could be better for
other enterprises which tend to have lesser geographical distributions
of data centers, is it so ?

Criticism
They donot cite the paper http://www.cs.rutgers.edu/~ricardob/papers/hotpower09.pdf
, which adopts a very similar approach of developing a cost aware
routing scheme. Unlike this paper they propose an optimization model
which can distribute queries according to two main schemes GreenDC
( route to green data centers) and EPrice ( route to cheaper data
centers ). I believe this paper provides a much simpler model than
this paper, but needs to be further analyzed to get a better
perspective.

This paper was really difficult to analyze and I think need more time
to read through it to analyze all the math and the electricity price
models.

[Andrew Ferguson]

Paper Title
"Cutting the electric bill for internet-scale systems"

Authors

Qureshi, Asfandyar, Weber, Rick, Balakrishnan, Hari, Guttag, John, and
Maggs, Bruce

Date
SIGCOMM 2009

Novel Idea

The authors suggest that operators of internet-scale systems take
advantage of their ability to route requests globally so that data
centers which have the cheapest energy costs take the most requests.
As a consequence, data centers with high energy costs (due to high,
fluctuating electrical prices) can be temporarily shut down to save
costs.

Main Results
The paper presents the results of a simulation in which a CDN with
locations in various electrical markets serves requests in order to
minimize a fluctuating cost of electricity. The traffic pattern is
based on a sample of traffic at Akamai public clusters, and the cost
of electricity is replayed from historical data. The simulation
suggests that a CDN with this setup could achieve substantial cost
savings.

Impact
Unknown.

Evidence
Detailed graphs present the case that energy prices across the U.S.
vary widely and asynchronously on a short time scale.

Prior Work
The authors do not mention any prior work.

Competitive Work
Unknown.

Reproducibility
The electricity prices are public information, so one could use that
data to try and recreate the analysis. However, one would need to get
a representative traffic sample in order to reproduce the simulation.

Question
Do other people think this is a good idea?

Criticism
I would have appreciated some theoretical analysis to complement the
empirical studies; a lot of work has been done on modeling electricity
prices, such as with stochastic processes. The authors did not address
how improvements to the electric grid might affect this work -- would
such improvements create a greater uniformity of prices? Figures 7b
and 9 do not look very Gaussian, which is what white noise would look
like -- this suggests that there are still correlations in that data
which could be exploited or modeled. On a practical level, don't data
center planners take local electricity pricing into account already?
Isn't that why Google and Microsoft have built data centers near The
Dalles? Why did the authors choose $5/MWh as their differential
threshold? Can we get a sensitivity analysis please? Given that
Akamai's CDN is sensitive to latency, as they admit, why did they
assume that non-latency-sensitive CDNs would build data centers in the
same way as Akamai?? Finally, it is expensive to do anything in NYC,
but that is also a population and internet hub, so companies might not
want to move from there.

Ideas for further work
How does this analysis apply to other countries? Would a routing
scheme optimized by an LP work better than the greedy scheme used in
the simulation?

[Marcelo Martins]

Paper Title "Cutting the Electric Bill for Internet-Scale Systems"

Author(s) Asfandyar Qureshi, Rick Weber, Hari Balkrishnan, John Guttag,
Bruce Maggs

Date SIGCOMM'09, August 2009

Novel Idea

The conventional approach to reducing energy costs has been to reduce
energy consumption. Qureshi et al. support the exploiting of request
routing and temporal and geographical variations of electricity for
diminishing the energy costs of a large data server. Requests from
clients are forwarded by routers to the data center located in the
cheapest area at a given time.

Main Result(s)

See (Evidence)

Impact

The cost aware request routing policy maps requests to locations where
energy is cheaper. Consequently, the energy cost of running large
Internet-scale systems is reduced.

Evidence

The authors carefully designed their argument of costs savings by a
thorough analysis of the electricity charging in various regions of the
US. Their findings include:

1) Servers and data centers accounted for 1.5% of US electricity
consumption in 2006
2) Regional Transmission Organizations (RTOs) determine the electricity
distribution over the grid for different regions of the US. Also, they
have different pricing policies.
3) There is a significant amount of day-to-day volatility on electricity
prices charged by providers
4) Rapid (hourly) price fluctuations reflects the momentary demand for
electricity
5) Prices in locations belonging to different RTOs are never highly
correlated and even those in the same region are not always correlated
6) Price differentials between two regions and its duration depend on
the time of day

Prior Work

There is an increasing interest in energy-proportional servers and
dynamic server provisioning techniques such as GreenFS and proposals
from eBay and Microsoft and have been explored.

Competitive work

None

Reproducibility

Mosts of the results were obtained through simulation. Given access to
the simulator or the specifics of its implementation or access to its
source code, plus the data provided by Akamai, most of the results could
be replicated.


Question/Future Work

1.) Can we estimate the threshold for a data center to participate in
demand-response programs without having large losses but also paying
lower prices?

Criticism

1.) Most of the results and estimations are based on a optimal hardware
savings model that probably cannot be achieved (0% idle, ). Current
hardware shows that for the actual energy parameters (65% idle, 1.7
PUE), the savings are much lower and only applicable to a few large
companies.

2.) The idea of routing server requests based on transitional pricing
and locality only games the energy system and does not reduce energy
consumption. Exploiting RTOs should not be taken as serious and relevant
as dropping actual power draw, since the former only takes advantage of
current conditions that might not exist in the future, while the latter
actually tries to promote a economical and environmental-friendly solution.

3.) The authors do not take into account that request routing can affect
the load distribution of a network. New QoS techniques or even backup
servers have to be allocated to respond to an unexpected flush of remote
requests due to pricing conditions.

[Dongbo Wang]

Paper Title: Cutting the Electric Bill for Internet-Scale Systems

Authors: Asfandyar Qureshi, Rick Weber, Hari Balakrishnan, John Guttag, and Bruce Maggs.

Date: 2009

Novel Idea: This paper is very interesting that it does not focus on how to make large scale systems more energy efficiency, but, from the practical point of view, try to cut the money wasted on electric energy by routing requests to servers that is located at low-energy-price districts. Based on the existing technology for routing requests, the requests can be routed to servers with lower energy unit price. It's true that the energy price varies in different location, so when the system has energy elasticity, it is feasible to save energy cost by migrating requests.

Main Result & Impact: the paper identifies the relevance of electricity price differentials to large distributed system and estimates the energy cost savings through simulation. The paper takes into account the bandwidth and performance. The paper argues that existing systems can save energy cost by at least two percent without losing too much performance and increasing bandwidth; the better the energy elasticity is, the more obvious the cost saving will be.

Evidence: the paper simulates and estimates the cost of a cost-aware routing scheme using 24-days real request traffic data from Akamai content distribution network. The results of the simulation reflects the effect of cost-aware routing scheme, but also shows that the cost is strongly related to the threshold of client-server distance.

Prior Work & Competitive Work: Many other works focus on increasing the energy elasticity. The improvement of energy elasticity will complement the work in this paper.

Reproducibility: I don't think we can reproduce the work in this paper.

Question & Criticism: none

2009/11/24 Rodrigo <rodrigo...@gmail.com>

[小柯]

Paper Title:    Cutting the Electric Bill for Internet-Scale Systems
Authors:        Asfandyar Qureshi

                    Rick Weber
                    Hari Balakrishnan
                    John Guttag
                    Bruce Maggs

Date:           2009

Novel Idea:
    This is necessarily an extremely new idea...a computer science paper talking about how to save money!
    Authors found that there's no correlation between electricity price and locations, which means that it would be more cheaper if most of the customers' requests could be routed to machines located in the cities or states that has lower electricity price.
    They present a method takes the electricity price into account to route the request and simulate the result to prove that this method works. (Saving money!)

Main Result:
    Based on a scheme taking electricity price into account, authors builds up a simulation model using Akamai traffic to show this scheme could reduce the electricity expense.

Impact:
    I am not sure if there would be any further research in this topic. A general research seeks for reduction of energy consumption.

Evidence:
    Authors show their observation on variation of electricity price in different location. Based on this observation and some analysis, they build up formulation to estimate electricity expense and using traffic collected from Akamai to establish the simulation model, which is later used for evaluation of their routing scheme.

Prior Work:

Competitive work:

Reproducibility:
    Yes. Just need to note any changes of electricity price.

Question:
    One assumption made by the author is the system is fully replicated, so any cluster could serve any of the request. Without this assumption, how is the effectiveness of this approach?

Criticism:
    There's a trade off between request latency and cheaper electricity bill.
    One could put all their servers in one location with the lowest electricity price. This might save much money; however, any place far away from this place would incur some latency.

Ideas for further work:

2009/11/23 Rodrigo <rodrigo...@gmail.com>

[Dan Rosenberg]

Paper Title
Cutting the Electric Bill for Internet-Scale Systems

Authors

Asfandyar Qureshi, et al.

Date
August, 2009

Novel Idea
The authors propose a request routing policy that is aware of energy costs, in
an attempt to save money for providers of large-scale distributed services.

Main Result/Impact
The authors convincingly demonstrate that there is at least some opportunity to
be had in reducing the cost of servicing requests by optimizing the purchase of
cheap electricity.

Evidence
The authors test their techniques in simulation, using a subset of Akamai's
usage statistics in conjunction with electricy market data on some of the
corresponding cities.  This simulation seems somewhat limited in scope, but the
results are strong.  Good charts.

Prior Work
This paper complements techniques used to minimize energy consumption.

Reproducibility
Without access to usage data, market data, and more details on implementation,
I would not be able to reproduce these experiments.

Criticism
The numbers for Figure 1 are a bit shaky - the authors admit they are based on
rough estimates.  Of the two major assumptions mentioned in this paper (1. open
market electricity is cheaper than long-term contracts and 2. variable energy
costs are a major expense for these systems), the first seems like an
assumption that may not hold.

Questions/Ideas for Further Work
The authors mention that applying this sort of solution to smaller scale
providers is future work.  I would also be curious as to whether widespread
adoption of a system like this might actually impact the electricity market
itself and introduce complications.  I especially like the idea of using these
techniques in conjunction with an environmental impact cost function.

[Juexin Wang]

Paper Title
Cutting the Electric Bill for Internet-Scale Systems

Author(s)

Asfandyar Qureshi
Rick Weber
Hari Balakrishnan
John Guttag
Bruce Maggs


Date

Aug, 2009

Novel Idea

By being cognizant of locational computation cost difference, large Internet scale companies can reduce electric cost without deploying new equipment nor specific technology. 

Use a novel way to collect traffic data.

Main Result(s)
This paper develops and analyzes a new method to reduce the energy costs of running large Internet-scale systems. It relies on electricity prices vary and  Large distributed systems already incorporate request routing and replication. Their results imply that existing systems may be able to save millions of dollars a year in electricity costs, by being cognizant of locational computation cost difference.


Impact
Companies can relocate their computing and data centers to save money on electricity costs. 

Evidence

- This paper provided us detailed data about electric price difference on hourly and geography.

- They use realistic workload -- Akamai, to understand the interaction of real workloads with electricity prices. They acquired a data set detailing traffic on Akamai’s infrastructure, covered 24 days worth of traffic on a large subset of Akamai’s servers, with a peak of over 2 million hits/sec. 
- By using data from the Akamai CDN as a representative real-world workload, they model the energy consumption in order to estimate by how much this idea can reduce energy cost.

Prior Work

N/A

Competitive work

Other cooling and energy saving system that really reduce the energy cost.

Reproducibility

No need to do that.

Question & Criticism

Could the Akamai workload represent the true load of internet transmission of all kinds of services and companies?

It's true that the electricity prices vary for different locations or time, but that also means this varies can be easily changed by prower providers.

And the routing will bring extra latency and transmission power cost. In all, it's not environment friendly but do more waste of energy. 

On Mon, Nov 23, 2009 at 6:55 PM, Rodrigo <rodrigo...@gmail.com> wrote:

Please post your reviews here.

--
J.W
Happy to receive ur message

[qiao xie]

A new method to cut the electricity bills using cost-aware requests routing policy.

The authors observe that electricity prices are quite different geographically and today's internet-scale web services are also geographically distributed.

 

The goal is to optimize the mapping of requests to geographically distributed servers to achieve a minimal cost of energy.

 In their schema, clients may be routed to distant servers in search of cheap energy. This idea implies a longer latency which might lead to bad user experience.

 I think the method should be applied to senarios that instant responses are not important. 

The paper modeled the enery consumption of a cluster linear to its utilization. They thought the CPU utilization is a good estimator for power usage, which might not be true according to the first paper. They performed simulations and estimated the savings.

In this paper, they focused on one factor of the cost model and developed a routing mechanism accordingly.

In the future, they plannded to introduce more factors such as bandwidth cost and weather differentials into the model to assist decision making.

I think performance issues should also be taken into account because if the users are unhappy companys will lose more money. 

2009/11/24, Rodrigo <rodrigo...@gmail.com>:

[Kevin Tierney]

Title: Cutting the Electric Bill for Internet-Scale Systems

Author(s): Qureshi, et. al.
Date: SIGCOMM 2009

Novel Idea
This paper proposes factoring the price of energy for data centers
based on the hourly wholesale price of power into routing
determinations.

Main Result(s)
Knowledge of where energy is cheaper at a given time of day allows the
authors to save money by moving computation away from data centers
during times of expensive power prices. The authors take into account
bandwidth costs by making a hard cap between data centers based on
data they received from Akamai, but also show how much money could be
saved if this cap were relaxed.

Impact
This paper could have a big impact on data center routing. It brings a
whole new idea into the objective function that can be used for
choosing data centers to serve clients.

Evidence
The authors studied 24 days worth of Akamai logs along with
synthesized data based on the information in the logs. They also
gathered price information for power in the areas around Akamai's
"public clusters".

Prior Work
There doesn't seem to be any, this paper lacks a prior work section.

Reproducibility
No, the algorithm for assigning clients to data centers is not all that clear

Question
How does this effect latency?

Criticism
The vagueness in the algorithm is disapointing, and further more, why
didn't they use an LP? Or did they, and we just don't know about it?
Also, instead of capping bandwidth, why not try to find some data on
pricing for those links and bring that into the objective function?
Then the whole thing could be globally optimized!

Although geography is given consideration, a study of the effect on
latency is omitted. It would be important to know whether Akamai would
be violating any SLAs through such a scheme.


Ideas for further work
Figures someone is already doing this with an LP- that's the first
thing that popped into my head when I saw that their calculation
functions could be well approximated with linear values...

On Mon, Nov 23, 2009 at 6:55 PM, Rodrigo <rodrigo...@gmail.com> wrote:

[Spiros E.]

The paper argues for considering the cost of energy in routing
requests in a distributed system. The goal of such a system is to
reduce the cost of energy consumption, which may cause overall energy
consumption to increase. The paper assumes a geographically
distributed system whose locations have energy prices with low
correlation.

Energy elasticity is discussed throughout the paper. The paper in
addition asserts that systems will in the future reach zero-idle
energy consumption, pointing to increased interest in the issue. This
is convenient since such an accomplishment would greatly increase the
efficacy of the paper's proposed techniques.

The paper does mention that the energy price differences between
geographic regions is sometimes non-zero. This would suggest that it's
always better to route traffic to the servers in a single geographic
location all the time (assuming a certain insensitivity to latency).
Or at least, that is a very simple heuristic one could use to route
traffic in a geographic region that will save a company money with
little effort.

The evaluation is entirely based on simulations run from historical
data. So though the paper has demonstrated that its ideas could lead
to savings in theory, no implementation of the proposed system exists.
This is odd, as the massive savings that the paper promises should
have been enough to entice an industrial partner (namely Akamai) to
try and implement the system.

In addition, the Akamai data that the evaluation is based on is (I
presume) not publically available. It's therefore impossible to
reproduce the findings of the paper. However, one could take data from
his own system and compare it to the same historical prices to see if
such a system would have led to savings.