MarketBased Resource Allocation for

1
Market-Based Resource Allocation for Utility
Data Centers Byde et al. HP Labs Taewon Hwang
2
Introduction ? Today, organizations must adapt to
their environments fast. ? IT became a key
determinant of a company's overall
effectiveness. ? To ensure that sufficient
resources are available to support growth, IT
planners have traditionally overbuilt and
underutilized the datacenter resources or
infrastructure. ? As a result, IT infrastructure
has evolved into a complex collection of legacy
systems that are hard to understand and manage in
a uniform fashion.
3
Introduction ? Traditionally, IT infrastructure
is static, requiring significant manual
intervention for each change. ? However, today's
enterprise is far more dependent on the
availability of IT service to conduct business
and demands a flexible datacenter architecture. ?
To address this issue, HP introduce Utility Data
Center (UDC), which allowed large scale,
heterogeneous IT infrastructure (servers,
storages, networks, security devices, etc.) to be
aggregated into a dynamically delivered resource
pool for applications.
4

• Introduction
• ? In this context (dynamic resource allocation),
  under-utilized resources can be redeployed
  wherever they are most needed and a service that
  is experiencing low load can relinquish resources
  (e.g. bandwidth or CPU cycles) to another service
  that is experiencing high load.
• ? Market-based resource allocation system has
  three high-level components.
• - Resource Market
• - Application Agent
• Business Agent

5
(No Transcript)
6
Market Mechanism ? In market-based approach,
different agents have different knowledge, goals
and preferences, but nonetheless must express
their desires in a common currency. ? How to
resolve the issue of who gets the resource? -
bidding ? Markets are at their most useful when
the participants are unwilling to reveal their
information directly. However, because the agents
are owned and operated by the UDC owner, they are
essentially cooperative.
7

• Market Mechanism
• ? Mechanism choices (general assumptions)
• Servers or other computing devices are the
  essential resources.
• 2. Resources are rented from the UDC in time
  periods.
• 3. The market is centralized.
• 4. It is not necessary to collect payments for
  resources.

8
Application Agent ? Glue between the business
level and the IT level. ? It should understand
both resource descriptions and service metric
specifications. Business Agent ? It should
analyze both SLA and strategic value.
9

• Simulations
• ? Market Mechanism (Simple Market)
• Find the bids of all agents for all servers A
  server is tradable if the highest bid for it is
  not that of the server currently in possession of
  it.
• 2. If there is at least one tradable server, find
  the server whose trade value is greatest, and
  effect the trade.
• 3. If a server was traded then start again from
  step 1.
• - This mechanism was used in most of the
  experiments.

10

• Simulations
• ? Market Mechanism (Globally Optimal Assignment
  Market)
• Get the bids bida(n) of each agent a for each
  number of servers it may receive, n, from 0 up to
  the total number of servers available.
• 2. Construct the set of all assignments (n1,,nA)
  of numbers of servers to agents choose the
  assignment (n1,,nA) for which total value is
  maximized
• 3. Assign na servers to agent a, for each a.

Value(n1,,nA)
11
Service Specification
12
Service Specification ? Cn is the sum of the
capacities of all servers allocated to processing
node n. Here, the time taken for the node to do
a unit of work (time per unit work) mn
(vn/Cn) ? mn is an un-parallelizable component
of the processing time that does not depend on
the set of resources allocated to the node,
whereas vn represents the part that is
parallelizable, and hence which will be affected
by adding resources to the node.
13
Service Level Agreements (SLAs) ? The UDC
operator is rewarded for completing jobs on time,
and pays penalties for delivering jobs late or
not at all.
14
Business Agent Algorithm ? The amount bid for
the resources that the agent actually received is
averaged over the lifetime of a service. ? In
order to accomplish such application, the authors
spread the predicted utility of all jobs in the
queue out over the period of time that is
predicted for the last job to exit the queue
the longest service time.
15
Application Agent Algorithm ? The agent
maintains an estimate of the average amount of
work per job for each processing node, (av
work)n, and an estimate (av length)n of the queue
length at each node. ? max(? min, (1/n1)) (av
length)n ?(Q length)n (1- ?)(av
length)n ? This update rule smoothes out the
observations of actual queue length over multiple
time steps, but ensures that (av length) is kept
up-to-date by placing a lower limit of ?min on
the weight given to new observations is not less
than ?min.
16
Application Agent Algorithm ? To estimate the
time at which job j, currently either being
processed or waiting to be processed at node n,
will be delivered (Q position)j (av work)n
workn,j (time per unit work) n (av length)m (av
work)m workm,j (time per unit work)m ? Where
(Q position)j is the number of jobs ahead of j in
the queue for processing node n, and workm,j
is the amount of work node m has to do on job j.

17
Application Agent Algorithm ? To allow the agent
to manage risk sensibly, the expected utility of
processing each job is evaluated not just with
respect to this delivery time, but with respect
to a cluster of times around this value, with
corresponding approximate likelihoods.
expected_utility(job,t) wj utility(job, sj
t) ? Scale factors sj (j 1,..,S) and
associated weights, wj. In simulations, arbitrary
values are assigned to these parameters.
18
Results ? Basic setup - It consists of two
statistically identical services (no gold or
sliver customers). - The resource pool for these
services typically consisted of 20 identical
servers. - A typical run lasted 1000 time
steps, and results regarding average performance
were taken over 100 such runs. - Each job was due
to be delivered 10 time steps after arriving in
the first queue, would pay 1 unit of money on
successful delivery, and would incur a fine of 1
if late.
19
Results ? Basic setup
0.88
The optimal static assignment is to give half the
servers to each service, in which case almost all
(94.0) jobs are completed on time
20
Results ? Basic setup - The average earnings
per job for the market controlled scenario is
0.94 (completing 97 of jobs on time). - The
average earnings per job for the symmetric static
allocation is 0.88 (completing 94 of jobs on
time). - The ability to re-provision resources
dynamically in response to business impact does
not lead to dramatic changes in the combined
effectiveness of the two services.
21
Results ? Market Volatility
The volatility of the allocations
The volatility of the job stream
Work Arrival rate 200
20 1200 3.3 3200
1.25
Demand fluctuates less, and hence so does the
allocation
22
Results
? Emergent Time Sharing
- The clear advantage of the dynamic allocation
mechanism is due to the fact that there is no
symmetric static allocation in this scenario.
23
Results
? Emergent Time Sharing
The market mechanism discovers the time-sharing
solution.
24

• Results
• ? Admission Control
• Work (5000), arrival rate (1) -gt highly
  fluctuating demand
• - Using the market-based resource allocator
  allows more services to be hosted, by increasing
  utilization of the underlying
• infrastructure.

25
Results ? Heterogeneous Value
2 4 6
8 10
26
Results ? Heterogeneous Value
Workload
4000
5000
- The market automatically makes the
necessary trade-offs to provision more resources
to application 2 than application 1.
6000
7000
27
Results ? Mechanism Comparison
Basic market
static
global
- For a fixed mechanism, using the true job
delivery times actually leads to worse
performance.
simulation global
In simulation, the agent has access to actual
delivery time
simulation
28
Conclusions ? The market mechanism never
performed worse than a static allocation, and
often performed significantly better. ? When
demand fluctuates highly, it was able to balance
application requirements so as to give earnings
that exceeded those from hosting fewer
services. ? However, In environment, market
mechanism choices have far less impact on
allocation value than bidding algorithm design
(need to be addressed in the future work).