Transcript RT-FT-Optimization (update thread)

Team 7: SUPER B RBAMAN
18-749: Fault-Tolerant Distributed Systems
Electrical &Computer
ENGINEERING
Team Members
Mike Seto
mseto@
Wee Ming
wmc@
Ian Kalinowski
igk@
Jeremy Ng
jwng@
http://www.ece.cmu.edu/~ece749/teams/team7/
or just
for ‘borbaman’
2
Baseline Application 1
SYNOPSIS:
–
–
–
–
–
Fault-tolerant, real-time multiplayer game
Inspired by Hudsonsoft’s Bomberman
Players interact with other players, and their actions affect the shared
environment
Players plant timed bombs which can destroy walls, players and other bombs.
Last player standing is the winner!
PLATFORM:
Middleware: Orbacus (CORBA/C++) on Linux
Graphics: NCURSES
Backend: MySQL
3
Baseline Application 2
COMPONENTS:
–
–
–
Front end: one client per player
Middle-tier game servers
Back end: database
STRUCTURE:
–
–
–
–
–
A client participates in a game with other clients
4 clients per game (more challenging and has real-time elements)
A client may only belong to one game
A server may support multiple games
Game does not start until 4 players have joined
4
Baseline Architecture
New Game
(4 Clients
each)
joinGame()
Game Server
(Server Object)
Server(s)
db_add_player()
Database
Create game
object
Clients
Active Game
(4 Clients
each)
update()
db_get_gameState()
db_set_gameState()
Game
Object(s)
Persistent
State of
Current
Games
(Bombs,
Players, etc)
5
Fail-Over Measurements – Fault-free
RTT vs. Number of Invocations
250000
RTT (us)
200000
150000
Database RTT
Client RTT
100000
50000
0
1
68
135 202 269 336 403 470 537 604 671 738 805 872 939 1006
Number of Invocations
6
Fault-Tolerance Goals
ORIGINAL FAULT-TOLERANT GOALS:
1) Preserves current game state under server failure
•
•
•
•
Coordinates of players, and player state
Bomb locations, timers
State of map
Score
2) Switch from failed server to another server within 1 second
3) Players who “drop” may rejoin game within 2 seconds
7
FT-Baseline Architecture
Stateless Servers
–
–
Two servers run on two machines with a single shared database
Passive replication:
• No distinction between “primary” and “backup” servers
• No checkpointing
–
–
–
Each server replica can receive and process client requests
But…clients only talk to one replica at any one time
Naming Service and Database system are single point of failure.
8
FT-Baseline Architecture
Guaranteeing Determinism
–
–
–
–
State is committed to the reliable database at every client
invocation
State is read from the DB before processing any requests, and
committed back to the DB after processing
Table locking per replica in the database insures atomic access per
game and guarantees determinism between our replicas
Transactions with non-increasing sequence numbers are discarded
Transaction processing
–
–
Database transactions are guaranteed atomic
Consistent state is achieved by having the servers read state from
the database before beginning any transaction
9
FT-Baseline Architecture
Server-1
Clients
1. update()
2. COMM failure
3. Retrieve
server list
X
Naming
Service
db_get_gameState()
db_set_gameState()
Database
db_get_gameState()
db_set_gameState()
4. retry()
Active Game
(4 Clients
each)
5. update()
Server-2
10
Mechanisms for Fail-Over
Failing-Over
1.
2.
3.
4.
Client detects failure by catching COMM/TRANSIENT exception
Client queries Naming Service for list of servers
Client connects to first available server in order listed in Naming Service
If this list is null, the client waits until a new server registers with the
naming service
11
Fail-Over Measurements – 16 Faults
RTT vs. Number of Invocations (with 16 Faults)
250000
RTT (us)
200000
150000
Client RTT
100000
50000
0
1
68
135 202 269 336 403 470 537 604 671 738 805 872 939 1006
Number of Invocations
12
Fail-Over Measurements – Breakdown with 16 Faults
13
Fail-Over Measurements
Problem:
• Average Fault-Free RTT: 14.7 ms
• Average Failure-Induced RTT: 78.8 ms
• Maximum Failure-Induced RTT: 1045.8 ms
Solution:
Too High!
Have servers pre-resolved by client, and have clients
pre-establish connections with working servers.
14
RT-FT-Baseline Architecture
What we tried:
–
–
Clients create a low-priority Update thread which contacts the Naming Service
at a regular interval, caches references of working servers, and attempts to preestablish connections.
This thread also performs maintenance on existing connections and repopulate
cache with new launched servers
What we expected:
X
15
RT-FT Optimization – Part 1
Before and after multi-threaded optimization
No Faults - Threaded
40000
40000
35000
35000
30000
30000
Invocation Time(us)
Invocation Time(us)
No Faults - Non Threaded
25000
20000
15000
25000
20000
15000
10000
10000
5000
5000
0
0
1
61 121 181 241 301 361 421 481 541 601 661 721 781 841 901 961
1
61 121 181 241 301 361 421 481 541 601 661 721 781 841 901 961
Invocation # (spaced every 50ms)
Invocation # (spaced every 50ms)
What went wrong?
16
Bounded “Real-Time” Fail-Over Measurements
Jitter:
–
–
–
Maximum Jitter BEFORE: 36 ms
Maximum Jitter AFTER: 176 ms
We have “improved” the jitter in our system by -389% !
RTT:
–
–
–
Average RTT BEFORE: 13 ms
Average RTT AFTER: 21 ms
We have “improved” the average RTT by -59% !
Why??
–
–
High overhead from the Update thread
Queried the Naming Service every 200 us!
• Oops….
17
RT-FT Optimization – Part 2
RTT with No Faults - 500ms
Multi-threaded Optimization
300000
250000
RTT (us)
200000
150000
100000
50000
0
1
63 125 187 249 311 373 435 497 559 621 683 745 807 869 931 993
Number of Invocations
Reduced the update period from 200 us to 500 ms
18
RT-FT Optimization – Part 2
RTT with 16 Faults - 1 min apart - 500 ms - Buggy
Multi-threaded Optimization
600000
13 spikes above 200 ms
500000
RTT (us)
400000
300000
200000
100000
0
1
1026 2051 3076 4101 5126 6151 7176 8201 9226 10251
Number of Invocations
With faults…but why the high periodic jitter?
19
RT-FT Optimization – Part 2
RTT with 16 Faults - 1 min apart - 500 ms - Fixed
Multi-threaded Optimization
600000
3 spikes above 200 ms
500000
RTT (us)
400000
300000
200000
100000
0
1
1035 2069 3103 4137 5171 6205 7239 8273 9307 10341
Number of Invocations
Bug discovered and fixed from analyzing results
20
RT-FT Fail-Over Measurements
Average RTT:
–
41 ms
Jitter:
–
–
Average Faulty Jitter: 81 ms
Maximum Jitter: 480 ms
Failover time:
–
–
Previous max: 210 ms
Current max: 230 ms
However, these numbers are not realistically useful because
- Cluster variability influences jitter considerably
- Measurements are environment dependent
% of Outliers before = 0.1286%
% of Outliers after = 0.0296%
21
RT-FT-Performance Strategy
LOAD BALANCING:
–
–
–
Load balancing is performed on the client-side
Client randomly selects initial server to connect to
Upon failure, client randomly connects to another alive server
MOTIVATION:
–
–
Take advantage of multiple servers
Explore the effects of spreading a single game across multiple servers
23
22
Performance Measurements
RTT - 16 clients (4 Games) - With Load Balancing
3500000
3500000
3000000
3000000
2500000
2500000
RTT (us)
RTT (us)
RTT - 16 clients (4 Games) - Without Load Balancing
2000000
1500000
2000000
1500000
1000000
1000000
500000
500000
0
0
1
164 327 490 653 816 979 1142 1305 1468 1631 1794 1957 2120 2283 2446
Number of Invocations
1
102 203 304 405 506 607 708 809 910 1011 1112 1213 1314 1415 1516
Number of Invocations
Load Balancing Worsens RTT Performance
23
Performance Measurements
Load balancing decreased performance
–
–
This is counter-intuitive
One single-threaded server should be slower than multiple singlethreaded servers
• Load balancing should have improved RTT since multiple servers could
service separate games simultaneously
Server code was not modified in implementing load balancing
–
–
Problem has to be with concurrent accesses to the database
This pointed us to a bug in the database table locking code
• Transactions and table locks were out of order, causing table locks to be
released prematurely
24
Average RTT (µs)
Performance Measurements
Load-Balancing with DB lock bug fixed
25
Performance Measurements
Corrected Load Balancing
–
Load balancing resulted in improved performance
• Non-balanced average RTT: 454 ms
• Balanced average RTT: 255 ms
26
Insights from Measurements
FT-Baseline
–
Can’t assume failover time is consistent or bounded
RT-FT-Optimization (update thread)
–
–
Reducing jitter resulted in increased average RTT
Scheduling the update thread too frequently results in increased jitter and
overhead
Load Balancing
–
Load balancing can easily be done incorrectly
• Spreading games across multiple machines does not necessarily improve
performance
• It can be difficult to select the right server to fail-over to
• Single shared resource can be the bottleneck
27
Open Issues
Let’s discuss some issues…
–
Newer cluster software would be nice!
• Newer MySQL with finer-grained locking, stored procedures, bug fixes, etc.
• Newer gcc needed for certain libraries
–
–
Clients can’t rejoin the game if the client crashes
Database server is a huge scalability bottleneck
If only we had more time…
–
–
–
–
GUI using real graphics instead of ASCII art
Login screen and lobby for clients to interact before game starts
Rankings for top ‘Borbamen’ based on some scoring system
Multiple maps, power-ups (e.g. drop more bombs, move faster, etc.)
28
Conclusions
What we’ve learned…
–
Stateless servers make server recovery relatively simple
• But this moves the performance bottleneck to the database
–
–
–
Testing to see that your system works is not good enough – looking at
performance measurements can also point out implementation bugs
Really hard to test performance on shared servers…
Testing can be fun when you have a fun project 
Accomplishments
–
–
–
Failover is really fast for non-loaded servers
We created a “real-time” application
BORBAMAN = CORBA+BOMBERMAN is really fun!
29
Conclusions
For the next time around…
–
–
–
Pick a different middleware that will let us run clients on machines outside of
the games cluster
Run our own SQL server that doesn’t crash during stress testing O:-)
Language interoperability (e.g. Java clients with C++ servers) could be cool
• Orbacus supposedly supports this
–
Store some state on the servers to reduce the load on the database
30
And the winner goes to …..
31
Finale!

No Questions Please.

GG EZ$$$ kthxbye
32
Appendix
RTT - No Faults - 500 ms
Multi-threaded Optimization
RTT - No Faults - 200 us
Multi-threaded Optimization
500000
500000
450000
450000
400000
400000
350000
350000
300000
300000
250000
250000
200000
200000
150000
150000
100000
100000
50000
50000
0
0
1
59 117 175 233 291 349 407 465 523 581 639 697 755 813 871 929 987
1
59 117 175 233 291 349 407 465 523 581 639 697 755 813 871 929 987
Varying Server Load Affects RTT
33
RT-FT Fail-Over Measurements
Breakdown of RTT (with 16 Faults)
Misc., 13181.85539,
17%
Normal Time,
37864.08211, 47%
Retry(), 28490.5,
36%
34