Douglas Thain and Miron Livny Computer Sciences Department University of Wisconsin-Madison Bypass: A tool for building distributed systems

Building distributed systems is hard.

Bypass makes building split execution systems easy. Bypass is to split execution systems as yacc is to compilers

Problem: Unfriendly Machines › Many systems can distribute your jobs to available machines scattered around the world. (rsh, Condor, Globus, etc...) › But... the machines you have access to may not be properly equipped to run your job.

Problem: Unfriendly Machines › An unfriendly machine…  allows you to login under some identity.  allows you to execute your program.  might not have your files or a shared file system!  might not have space for your output!  might be a different architecture or OS! (If you want to use a lot of machines, you can’t be picky!)

home machine foreign machine HELP! core dumped

Solution: Split Execution › General strategy:  An agent process traps some of the application's standard library calls.  Some of the calls can be executed at the foreign machine.  Some of the calls are sent via RPC back to the home machine.  A shadow process at the home machine executes the RPCs and sends the results back to the agent.

Solution: Split Execution Shadow Kernel Local system calls Home Machine Agent Application Kernel Local system calls Trapped system calls Foreign Machine Remote system calls

Split Execution is an Open Research Topic › We want to explore many possibilities:  Foreign machine could be partially friendly – has some needed resources, but not all.  Data may be buffered and cached at both the agent and the shadow.  What procedure calls to trap depends on the application and the services needed.  Some procedure calls could be routed to third parties such as file servers.  …

Problem: Split Execution is Hard › One example of many: Trapping stat()  Different data types: struct stat, struct stat64 Depending on system, integer elements are 2->8 bytes  Multiple entry points: stat, _stat, __libc_stat  Surprises: #define stat(a,b) _fxstat(VERSION,a,b)

Solution: Bypass › Bypass takes a specification of a split execution system and produces a matched shadow and agent. › Bypass hides all of the ugly details of trapping, type conversion, and RPCs. › Bypass lets you:  split any dynamically-linked application.  transparently use heterogeneous systems.  trap calls with minimal overhead.  control execution paths with plain C++.

home machine foreign machine Just like home!

Bypass Language › Declare what procedures to trap in C++ › Annotate pointer types with data flow.  Direction: in, out, or in out  Binary data: give expression yielding the number of bytes to send/receive. › Give two function bodies:  agent_action  shadow_action

ssize_t write ( int fd, in "length" const void *data, size_t length ) agent_action {{ if( fd==1 ) { return bypass_shadow_write(fd,data,length); } else { return write(fd,data,length); } }} shadow_action {{ printf("remote data: %s", data ); }} ;

Agent Action › Any arbitrary C++ code. › When the program invokes write(), the agent_action is executed at the home machine. › Within the agent_action:  write() - Invoke the original write() at the foreign machine.  bypass_shadow_write() - Invoke the shadow_action via RPC.

Shadow Action › Any arbitrary C++ code. › If the agent decides to invoke the RPC to the shadow, the shadow_action is executed at the home machine. › Within the shadow_action:  write() - Invoke write() at the home machine.

Using Bypass › Run "bypass" to read the specification and produce C++ source code: % bypass -agent -shadow simple.bypass › The shadow is compiled into a plain executable. › The agent is compiled into a shared library.

Using Bypass › The dynamic linker is used to force the agent into an executable at run-time: setenv LD_PRELOAD simple_agent.so › Procedure calls are “trapped” merely by putting the agent first in the link list. › This method can be used on any dynamically- linked program: tcsh, netscape, emacs…

Example Application: Complete Remote I/O › Trap all the standard I/O calls, and send them to the home machine unmodified: open(in string char *path, int flags, int mode); close(int fd); int read(int fd, out “length” void *data, int length ); int write(int fd, in “length” void *data, int length ); int lseek(int fd, off_t offset, int whence );

Complete Remote I/O Shadow Kernel open, close, read, write, lseek Home Machine Agent Application Kernel all other calls Trapped system calls Foreign Machine open, close read, write, lseek

Example Application: Remote Console › Trap only read and write, and send operations on standard files back to a single shadow process. int read( int fd, in “length” void *data, int length ) agent_action {{ if( fd<3 ) { bypass_remote_read( fd, data,length ); } else { return read(fd,data,length); } }};

Remote Console Shadow Kernel read, write Home Machine Agent Application Kernel all other calls Trapped system calls Foreign Machine Agent Application Kernel all other calls Trapped system calls Foreign Machine Agent Application Kernel all other calls Trapped system calls Foreign Machine Standard I/O reads and writes

Example Application: Attach New Filesystem › Trap standard I/O calls and replace them with calls to a user-level filesystem library, such as Globus GASS. int open( in string const char *path, int flags, int mode ) agent_action {{ return globus_gass_open( path, flags, mode ); }}; int close( int fd ) agent_action {{ return globus_gass_close( fd ); }};

Attach New Filesystem Agent Application Kernel all other calls Trapped system calls Foreign Machine Globus Library more system calls open close THE GRID

Bypass can be used by Real Users! › Bypass works on unmodified executables.  (Real Users are not willing/able to rewrite/recompile their programs.) › Bypass requires no special privileges.  (Real Users do not have the root password) › Thus, Bypass allows a Real User to make good use of a remote cluster without begging the administrator to configure it to his/her needs.

Performance › Overhead of trapping a system call is very small: 1-4 us  The "trapping mechanism" simply interposes a few extra function calls.  Small compared to the expense of a real system call (about 10-70us) › Remote procedure calls are, as expected, much slower: about 1 ms under the best conditions.

Related Work › “Classic” RPC and XDR:  Define standard integer sizes, endianness, etc.  Start by defining external protocol, then produce programming interface which is not always convenient: struct read_results * read_1( int fd, int length );

Related Work › Bypass:  We are stuck with existing interfaces, so annotate them to produce a protocol: int read( int fd, out “length” void *data, int length );  Do “best effort” conversion to/from external data format: off_t is 4 bytes on some platforms, 8 bytes on others. A conversion might fail!  Define canonical values for source-level symbols: O_CREAT has different values on Linux and Solaris!

Related Work › Hunt and Brubacher, “Detours”  Trap library calls on NT using binary rewriting – can be applied to any executable.  Make original procedure available through special “trampoline” call.  Bypass leaves the original entry point intact, so subroutines need not be re-written to use the trampoline.

Related Work › Alexandrov, et al., “UFO”  Use a kernel-level facility to trap all of a process’ system calls and translate some of them into WWW operations.  The kernel mechanism is secure and can be applied to any process.  But… it has a high (7x) trapping overhead and cannot be applied to procedures that are not true system calls.

Related Work › Bypass:  Trapping overhead is very small and can be performed on procedures that are not necessarily system calls.  But… can only be applied to dynamically- linked executables, and is not suitable as a security mechanism.

Related/Future Work › A complete remote execution system needs both methods:  The program owner provides a lightweight mechanism for creating a correct split execution environment.  The machine owner provides a heavyweight mechanism to defend itself from a (possibly) malicious program.

Complete System Shadow Kernel open, close, read, write, lseek Home Machine Agent Application Kernel all other calls Foreign Machine open, close read, write, lseek Sandbox

Future Work › Multiple agents applied to one application  How to select and invoke the correct agent action? › Signal handling  Flow of control is backwards. › Other implementations  Binary rewriting.  Build specialized linker that understands multiple definitions of symbols.

Further Questions? › Douglas Thain  › Miron Livny  › Bypass Web Page  › Questions now?