1) FlexSC implements efficient and flexible system calls for the multicore era by introducing exception-less system calls and syscall threads. 2) Results show MySQL throughput increased up to 40% and latency reduced 30%, and Apache throughput increased up to 115% and latency reduced 50%. 3) Synchronous system calls are expensive due to direct mode switching costs and indirect processor structure pollution costs, which exception-less system calls help reduce.

1. FlexSC
Flexible System Call Scheduling with
Exception-Less System Calls
Livio Soares and Michael Stumm
University of Toronto

2. Motivation
The synchronous system call interface is a
legacy from the single core era
Expensive! Costs are:
➔ direct: mode-switch
➔ indirect: processor
structure pollution
FlexSC implements efficient and flexible
system calls for the multicore era
2

3. FlexSC overview
Two contributions: FlexSC and FlexSC-Threads
Results in:
1) MySQL throughput increase of up to 40%
and latency reduction of 30%
2) Apache throughput increase of up to 115%
and latency reduction of 50% 3

4. Performance impact of synchronous syscalls
➔ Xalan from SPEC CPU 2006
➔ Virtually no time in the OS
➔ Linux on Intel Core i7 (Nehalem)
➔ Injected exceptions with varying frequencies
➔ Direct: emulate null system call
Direct
➔ Indirect: emulate “write()” system call
Indirect
➔ Measured only user-mode time
➔ Kernel time ignored
Ideally, user-mode performance is unaltered
4

5. Degradation due to sync. syscalls
Degradation (lower is faster)
Xalan (SPEC CPU 2006)
70%
60%
Apache Indirect
50% Direct
40% MySQL
30%
20%
10%
0%
1K 2K 5K 10K 20K 50K 100K
user-mode instructions between exceptions
(log scale)
System calls can half processor efficiency;
indirect cause is major contributor
5

6. Processor state pollution
➔ Key source of performance impact
➔ On a Linux write() call:
rd
➔ up to 2/3 of the L1 data cache and data
TLB are evicted
➔ Kernel performance equally affected
➔ Processor efficiency for OS code is also cut
in half
6

7. Synchronous system calls are expensive
User
Kernel
Traditional system calls are synchronous
and use exceptions to cross domains
7

8. Alternative: side-step the boundary
User
Kernel
Exception-less syscalls remove synchronicity
by decoupling invocation from execution
8

9. Benefits of exception-less system calls
➔ Significantly reduce direct costs
➔ Fewer mode switches
User
➔ Allow for batching
Kernel
➔ Reduce indirect costs
➔ Allow for dynamic multicore specialization
➔ Further reduce direct and indirect costs
9

10. Exception-less interface: syscall page
write(fd, buf, 4096);
entry = free_syscall_entry();
/* write syscall */
entry->syscall = 1;
entry->num_args = 3;
entry->args[0] = fd;
entry->args[1] = buf;
entry->args[2] = 4096;
entry->status = SUBMIT;
SUBMIT
while (entry->status != DONE)
DONE
do_something_else();
return entry->return_code;
10

11. Exception-less interface: syscall page
write(fd, buf, 4096);
entry = free_syscall_entry();
/* write syscall */
entry->syscall = 1;
entry->num_args = 3;
entry->args[0] = fd;
entry->args[1] = buf;
entry->args[2] = 4096; SUBMIT
entry->status = SUBMIT;
SUBMIT
while (entry->status != DONE)
DONE
do_something_else();
return entry->return_code;
11

12. Exception-less interface: syscall page
write(fd, buf, 4096);
entry = free_syscall_entry();
/* write syscall */
entry->syscall = 1;
entry->num_args = 3;
entry->args[0] = fd;
entry->args[1] = buf;
entry->args[2] = 4096; DONE
entry->status = SUBMIT;
SUBMIT
while (entry->status != DONE)
DONE
do_something_else();
return entry->return_code;
12

13. Syscall threads
➔ Kernel-only threads
➔ Part of application process
➔ Execute requests from syscall page
➔ Schedulable on a per-core basis
13

14. System call batching
Request as many system calls as possible
Switch to kernel-mode
Start executing all posted system calls
Avoids direct and indirect costs,
even on a single core 14

15. Dynamic multicore specialization
FlexSC makes specializing cores simple
Dynamically adapts to workload needs
15

16. What programs can benefit from FlexSC?
Event-driven servers
(e.g., memcached, nginx webserver)
➔ Use asynchoronous calls, similar to FlexSC
➔ Can use FlexSC directly
➔ Mix sync and exception-less system calls
Multi-threaded servers: FlexSC-Threads
➔ Thread library, compatible with Pthreads
➔ No changes to app. code or recompilation required
➔ Transparently converts legacy syscalls into
exception-less ones
16

17. FlexSC-Threads library
➔ Hybrid (M-on-N) threading model
➔ One kernel visible thread per core
➔ Many user threads per kernel-visible thread
➔ Redirects system calls (libc wrappers)
➔ Posts exception-less syscall to syscall page
➔ Switches to other user-level thread
➔ Resumes thread upon syscall completion
Benefits of exception-less syscalls
while maintaining sequential syscall interface
17

18. FlexSC-Threads in action
User
18

19. FlexSC-Threads in action
On a syscall:
Post request to system call page
Block user-level thread
19

20. FlexSC-Threads in action
Kernel
On a syscall:
Post request to system call page
Block user-level thread
Switch to next ready thread 20

21. FlexSC-Threads in action
User
Kernel
If all user-level threads become blocked:
1) enter kernel
2) wait for completion of at least 1 syscall
21

22. Evaluation
➔ Linux 2.6.33
➔ Nehalem (Core i7) server, 2.3GHz
➔ 4 cores on a chip
➔ Clients connected on 1 Gbps network
➔ Workloads
➔ Sysbench on MySQL (80% user, 20% kernel)
➔ ApacheBench on Apache (50% user, 50% kernel)
➔ Default Linux NTPL (“sync”) vs.
sync
FlexSC-Threads (“flexsc”)
flexsc
22

23. Sysbench: “OLTP” on MySQL (1 core)
500
400
(requests/sec.)
Throughput
300
15% improvement
200
flexsc
100
sync
0
0 50 100 150 200 250 300
Request Concurrency
23

24. Sysbench: “OLTP” on MySQL (4 cores)
1,000
800
(requests/sec.)
Throughput
600
40% improvement
400
200 flexsc
sync
0
0 50 100 150 200 250 300
Request Concurrency
24

25. MySQL latency per client request
256 connections
1900
1,000
900 95th
800 percentile
Latency (ms)
700 average
600
500
400
300
200
100
0
sync flexsc sync flexsc sync flexsc
1 core 2 cores 4 cores
Up to 30% reduction of average
request latencies
25

26. MySQL processor metrics
SysBench (4 cores)
1.4
1.2
Relative Performance
User Kernel
1
(flexsc/sync)
0.8
0.6
0.4
0.2
0
L3 d-cache TLB IPC L2 i-cache Branch
IPC L2 i-cache Branch L3 d-cache TLB
Performance improvements consequence of
more efficient processor execution
26

27. ApacheBench throughput (1 core)
45,000
40,000 flexsc
35,000 sync
(requests/sec.)
Throughput
30,000
25,000
20,000
15,000 80-90% improvement
10,000
5,000
0
0 200 400 600 800 1000
Request Concurrency
27

28. ApacheBench throughput (4 cores)
45,000
40,000
35,000
(requests/sec.)
115% improvement
Throughput
30,000
25,000
20,000
15,000
10,000 flexsc
5,000 sync
0
0 200 400 600 800 1000
Request Concurrency
28

29. Apache latency per client request
256 concurrent requests
238
30
99th
25 percentile
Latency (ms)
20 average
15
10
5
0
sync flexsc sync flexsc sync flexsc
1 core 2 cores 4 cores
Up to 50% reduction of average
request latencies
29

30. Apache processor metrics
Apache (1 core)
2
Relative Performance
1.5
(flexsc/sync)
User Kernel
1
0.5
0
L3 d-cache TLB IPC L2 i-cache Branch
IPC L2 i-cache Branch L3 d-cache TLB
Processor efficiency doubles for kernel
and user-mode execution
30

31. Discussion
➔ New OS architecture not necessary
➔ Exception-less syscalls can coexist with legacy ones
➔ Foundation for non-blocking system calls
➔ select() / poll() in user-space
➔ Interesting case of non-blocking free()
➔ Multicore ultra -specialization
➔ TCP Servers (Rutgers; Iftode et.al), FS Servers
➔ Single-ISA asymmetric cores
➔ OS-friendly cores (HP Labs; Mogul et. al)
31

32. Concluding Remarks
➔ System calls degrade server performance
➔ Processor pollution is inherent to synchronous
system calls
➔ Exception-less syscalls
➔ Flexible and efficient system call execution
➔ FlexSC-Threads
➔ Leverages exception-less syscalls
➔ No modifications to multi-threaded applications
➔ Throughput & latency gains
➔ 2x throughput improvement for Apache and BIND
➔ 1.4x throughput improvement for MySQL 32

33. FlexSC
Flexible System Call Scheduling with
Exception-Less System Calls
Livio Soares and Michael Stumm
University of Toronto