S3M (performance Mismatch in I/O Size, ParalleliSm and Sequentiality)

Haochen He, Erci Xu, Shanshan Li, Zhouyang Jia, Si Zheng, Yue Yu, Jun Ma, and Xiangke Liao. "When Database Meets New Storage Device: Understanding and Exposing Performance Mismatches via Configurations." PVLDB, 16(7). (TO APPEAR)

confSysTaint is the core of S3M, confSysTaint is based on LLVM IR, it analyzes the control and data dependency between configuration variable and specific syscalls. An example shows how it works:

Target syscalls

In section 4.1 Test Input Generation, we mention that S3M focuses on four series of syscalls, theay are obtained by our manual investigation on every Linux syscall (335 found in kernel version 5.4.0) by reading the official manual, followed by cross-checking. The filtered out 21 syscalls that may affect I/O size, parallelism, and sequentiality:

read series	write series	sync series	thread series
read pread64 readv preadv preadv2 io_submit io_getevents madvise open mmap	write pwrite64 writev pwritev pwritev2 io_submit io_getevents madvise open mmap	fsync fdatasync syncfs sync_file_range fcntl	clone(pthread_create) fork

Scenarios that confSysTaint supports

Data flow

• basic LLVM "Use" support
  
• Field sensitive analysis
  
• Inter-procedure (with pointer)
  
  
• Our extended data-flow (phi-node)
  
  • How to formaly determine if a phi-node will be tainted
    Given a phiNode like:

       phi i32 [ %5, %bb1.i ], [ 0, %bb1 ]
                  pre_node      pre_node2
    

    we check if:

Control flow

Formaly define how the control flow:

• Control Dependency: A block Y is control dependent on block X if and only if: Y post-dominates at least one but not all successors of X.
  • Transitivity：if A control dependent on B, B control dependent on C, then A control dependent on C.

An example, where the yellow square indicats the complicated code structures that motivate the use of the formal definition.

Usage

Dependency

• llvm-10.0.0
• gllvm

Build

cd tainter
cmake -DCMAKE_CXX_COMPILER=/usr/bin/clang++-10 -DCMAKE_C_COMPILER=/usr/bin/clang-10 -DLLVM_DIR=/usr/lib/llvm-10/cmake . 
make

Run

cd test/demo
../../tainter test.bc test-var.txt

For real DBMS, use gllvm to obtain the .bc file (e.g., mysqld.bc).

Check results

cat test-records.dat

Example result

Tainted Functions (group by Caller-Functions): 

		Clone_Handle::open_file <------------ func-1 of "srv_unix_file_flush_method"
				Clone_Task_Manager::set_error ----- Tainted Function.

		Clone_Snapshot::update_block_size <-- func-2 of "srv_unix_file_flush_method"
				os_event_set -------------------\ 
				pfs_unlock_mutex_v1              |_ Tainted Function.
				sync_array_object_signalled      |
				ut_dbg_assertion_failed --------/

		Double_write::sync_page_flush <------ func-3 of "srv_unix_file_flush_method"
				__clang_call_terminate ---------\ 
				buf_page_io_complete             |-- Tainted Function.
				fil_flush ----------------------/
        
    ...