hardware counter driven on-the-fly request signatures
DESCRIPTION
Hardware Counter Driven On-the-Fly Request Signatures. Kai Shen Ming Zhong Sandhya Dwarkadas Chuanpeng Li Christopher Stewart Xiao Zhang University of Rochester. Motivation. Hardware counters on modern processors: - PowerPoint PPT PresentationTRANSCRIPT
Hardware Counter Driven On-the-Fly Request Signatures
Kai Shen Ming Zhong Sandhya Dwarkadas
Chuanpeng Li Christopher Stewart Xiao Zhang
University of Rochester
04/19/23 ASPLOS 2008 2
Motivation
Hardware counters on modern processors: instruction mix, rate of execution, branch prediction
accuracy, memory access behavior
Operating system utilization of hardware counter metrics
Advantages as fine-grain workload signatures: application-transparency compared to application
statistics consistent availability compared to OS software statistics free fine-grain counter maintenance compared to
software statistics in general
04/19/23 ASPLOS 2008 3
On-the-Fly Request Signatures
Identifying requests for server workloads On-the-fly: identify a request while it still executes Utilizations:
Predicting request properties to guide OS adaptations Classifying requests on-the-fly to detect anomalies
0 20 40 60 80 100
4
6
8
10
x 10-4
Cumulative request execution (in millisec)
Flo
atin
g po
int
ops
per
CP
U c
ycle
TPC-H Q4TPC-H Q3TPC-H Q13
04/19/23 ASPLOS 2008 4
Challenges
Hardware metrics as workload signatures in server system environments
fluctuating concurrency and frequent context switches⇒ unstable hardware execution characteristics
requests are fine-grain workload units
Tracking request contexts within the OS on-the-fly transparent to applications
04/19/23 ASPLOS 2008 5
Hardware Metrics As Request Signatures:Choosing Normalization Base
Acquiring stable metrics as request executes: time-normalized metrics: divide by elapsed CPU cycles progress-normalized metrics: divide by retired
instructions
Finding: time-normalization for “time duration”-style metrics
(e.g., trace cache deliver mode)
0
0.1
0.2
0.3
0.4
0.5
Me
tric
-re
qu
est
-co
rre
latio
n
Non-halt CPU ticks
L2 misses (retired ins.)
L2 references
Load/store micro-ops
Trace cache deliver mode
Time-normalizedProgress-normalized
04/19/23 ASPLOS 2008 6
Hardware Metrics As Request Signatures:Choosing Effective Metrics
Environmental dynamics: concurrent request execution in server environments hardware resource-sharing – multi-threading and multi-core
Example metrics that are significantly affected:
0
0.1
0.2
0.3
0.4
0.5
Me
tric
-re
qu
est
-co
rre
latio
n
Load/store micro-ops
L2 misses
Data TLB pagewalk misses
Trace cache misses
L1 misses (retired ins.)
Serial executionConcurrent executionHyper-threading
04/19/23 ASPLOS 2008 7
Hardware Metrics As Request Signatures
Metric effectiveness across different applications inconsistent (e.g., floating-point ops very useful for
some but useless for others)⇒ Disappointing result: difficult to find a small set of
universally effective metrics
Require application-specific calibration
04/19/23 ASPLOS 2008 8
OS Support of Request Context Tracking
On-the-fly transparent tracking of request contexts Resource containers [Banga et al.’99] – not application-
transparent Magpie [Barham et al.’04] – not on-the-fly
High-level guidance: component activities reachable through control or data flows
are semantically related, and thus likely part of one request
One case: propagate request context through message passing
tag messages with senders’ request context IDs handle asynchronous messages, clarify message boundaries in
stream-based communications
04/19/23 ASPLOS 2008 9
Example of Request Context Propagation
Multi-tier RUBiS web server application
components database
Entirely at the OS
transparent to application
TomcatServelet
JbossInvoker
Jboss RMI-Disc
RUBiSQueryHome
RUBiSItemHome
MySQLDatabase
Request
... ... ... ... ... ... ... ...
... ... ... ... ... ... ... ...
Tim
e
04/19/23 ASPLOS 2008 10
Signature-driven Request Identification
Request identification: maintain a bank of recent past requests signature is a vector of metric statistics match each new request with banked requests on-the-
fly
Property inference: infer the property of new request using the property of
matched past request
04/19/23 ASPLOS 2008 11
Prototype
Platform Linux 2.6.10/Intel Xeon processors with hyper-threading
Overhead (not yet optimized):
0
0.5%
1.0%
1.5%
2.0%
Se
rve
r th
rou
gh
pu
t de
gra
da
tion
Metric collect/10msMetric collect/1msCollect/1ms with request identification
04/19/23 ASPLOS 2008 12
Evaluation Results:Accuracy of Predicting Request CPU Time
1 2 3 4 5 6 7 8 9 10
20%
40%
60%
80%
100%
Cumulative execution (in millisec)
Mea
n pr
edic
tion
erro
r
TPC-H
Running averageHardware counter
1 2 3 4 5 6 7 8 9 10
20%
40%
60%
80%
100%
Cumulative execution (in millisec)
Mea
n pr
edic
tion
erro
r
Comparison base (running average): the average properties of recent past requests to predict future requests
04/19/23 ASPLOS 2008 13
Utilization:Shortest-Job-First Scheduling
15-27% shorter response time than running average perform similar to oracle
3.1 3.2 3.3 3.4 3.50
2
4
6
Request rate (in reqs/sec)
Mea
n r
esp
ons
e tim
e (in
se
c) TPC-H
56 58 60 620
200
400
600
800
Request rate (in reqs/sec)
Mea
n r
esp
ons
e tim
e (in
mill
isec
)
RUBiS
Running averageHardware counterOracle
04/19/23 ASPLOS 2008 14
Utilization:Request Classification and Anomaly Detection
Dots are normal TPC-H requests
Circles are anomalies (SQL injection attacks)
10-ms cumulative metrics
0 2 4 6 8
x 10-3
0
1
2
3
4
5
6
7
8
x 10-3
Trace cache misses per m-instruction
Flo
atin
g p
oin
t o
ps
pe
rm
04/19/23 ASPLOS 2008 15
Related Work
Other uses of hardware counters phase detection [Dhodapkar&Smith’02, Sherwood et al.’03] behavior prediction [Duesterwald et al.’03, Bulpin&Pratt’05] anomaly tracking [Sweeney et al.’04]⇒ we handle challenges due to dynamic server environments
Request characterization using system software metrics tracking request/response [Aguilera et al.’03] request modeling [Barham et al.’04] failure diagnosis [Chen et al.’04]⇒ hardware metrics have unique advantages: consistent
availability, free fine-grain counter maintenance
First to realize on-the-fly request signatures for server workloads.
04/19/23 ASPLOS 2008 16
Conclusion
Our contributions: investigate the effectiveness of hardware counter
metrics as request signatures in dynamic server environments
propose OS mechanism to support on-the-fly request context tracking and adaptation
demonstrate the effectiveness of request signature-enabled on-the-fly OS exploitations
High-level takeaway: OS exploitation of hardware metrics to improve
performance and dependability [HotOS’07]