ActiveSLA: A Profit-Oriented
Admission Control
Framework for Database-as-a-Service
Providers
Pengcheng Xiong (Georgia Tech); Yun Chi (NEC Labs America); Shenghuo Zhu
(NEC Labs America); Junichi Tatemura (NEC Labs America); Calton Pu (Georgia
Tech); Hakan Hacigumus (NEC Labs America)
1
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
2
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
3
Database-as-a-service (DaaS)

DaaS provider consolidates multiple clients
in shared infrastructures (multi-tenancy)
greater economies of scale
 fixed cost distribution


Problem: system overload due to
unpredictable and more bursty workloads
dynamic provisioning,
 queuing and scheduling, and
 admission control

4
Admission control related work


Macro level (feedback based): keep the mean
query execution time at a specific level by tuning
the best multiple programming level (MPL) for a
given workload, e.g., ICDE2006
Micro level (query-by-query based): estimate
every single query’s execution time by query type
and query mix, e.g., WWW2004, ICDE2010
None of them has well addressed the problem
to directly maximize DaaS provider’s profits by
satisfying different SLAs for their clients!
5
First issue

Merely estimating the query execution time is
not enough to make profit-oriented decisions.
We need to know the probabilities of a query
meeting and missing its deadline.
Accept!
?
6
Second issue

We may have to make different admission
control decisions even when the queries have
the same deadline and the same probability of
meeting the deadline due to different SLAs.
7
System architecture of ActiveSLA
8
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
9
Prediction module design
 What
kind of models to use?
 The
model selection between linear and
nonlinear models, between regression
and classification models
 What
features to use?
 The
rich set of features for DaaS
providers
10
Model selection

Linear vs. Nonlinear


The execution time of a query depends on many factors in a
non-linear fashion, i.e., isolation levels and available buffer
size
Regression vs. Classification

From the machine learning point of view, a direct model of
classification usually outperforms a two-step regression based
approach.
11
Feature collection


Query Type and Mix (TYPE, Q-Cop, ActiveSLA)
Query Features (ActiveSLA)


E.g., the estimated number of sequential I/O
Database and System Conditions (ActiveSLA)




Buffer cache: the fraction of pages of each table that are
currently in the database buffer pool.
System cache: the fraction of pages of each table that are
currently in the operating system cache.
Transaction isolation level: Read Committed(FALSE) or
Serializable(TRUE).
CPU, memory, and disk status: the current status of CPU,
memory, and disk in the operating system.
12
Summary
13
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
14
Prediction module evaluation
 Query
Sets with PostgreSQL server
TPC-W1 (browsing queries)
 TPC-W2 (mixture of browsing and administrative
queries)
 TPC-W3 (mixture of browsing, administrative, and
updating queries)

 Prediction
error
False positive False negative
15
Total number
Prediction module evaluation
For different query sets(TPC-W2,3), different SLA settings (30s,
45s, 60s), we observe a steady decrease of errors when we use
non-linear model, classification model, and include more features
Details on the Machine Learning Model
Positive value->more likely to miss deadline
Negative value->unlikely to miss deadline
CPU waiting% for IO
Q10 is the update query
that needs exclusive lock
Lots of lock
contention
Very likely to
miss deadline
Details on the Machine Learning Model
Query Plan B
Query Plan A
Overhead and feature sensitivity

Overhead



Training overhead. 72ms to build an initial model by using
12,000 samples.
Evaluation overhead. 8ms
Feature sensitivity


19
The more features, the better
The gain by using more
features is less than the gain by
using a better model.
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
20
Decision module design
General SLA and PDF
of query execution time
Step-wise SLA and PDF
of query execution time
The probability that
the query can
meet deadline
21
Multiple Query Decision



Admitting q into the database server may slow down
the execution of other queries that are currently
running in the server and make them miss deadline.
Admitting q will consume system resources and change
the system status. This may result in the rejection of the
next query, which may otherwise be admitted and bring
in a higher profit.
Model this as opportunity cost o.
22
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusions
23
Decision module evaluation
 Result
with stationary workload (static
Poisson arrival rate)
 Result with non-stationary workload
(dynamic Poisson arrival rate
according to 1998 World Cup Trace)
 Single
SLA
 Multiple SLAs(service differentiation)
Result with stationary workload
Result with non-stationary workload
Single SLA summary
Multiple SLAs summary
Overview
 Motivation
 Prediction
module design
 Prediction module evaluation
 Decision module design
 Decision module evaluation
 Conclusion
28
Conclusion

We proposed a framework, ActiveSLA, for
admission control in cloud database systems.



Prediction module to predict the possibility that a query can
meet/miss deadline.
Decision module to make the profit-oriented decision.
Future work



Improve the inaccuracy for the query features such as the
number of sequential I/O due to the incorrect statistics and
cardinality estimates of a query execution plan.
Extend our prediction module by including the level of
replication as one of the system variables.
Extend our ActiveSLA to deal with different types of
database systems to manage data and serve queries, e.g.,
NoSQL databases.
29
Thank you!
Pengcheng Xiong is now on the job market
E-mail: xiong@gatech.edu
30