Scribd
Upload
56 views

Lock Elision Pthreads and MySQL

The document discusses lock elision, which is a technique supported in recent versions of glibc that allows certain lock acquisitions to be performed without actually acquiring the lock in cases where no contention is detected. The author is curious if this technique provides benefits for databases like MySQL and RocksDB that can experience contention on read-write locks when reads are frequent. They also note that documentation for glibc threading features like this can be limited, and provide a helpful summary from an expert on the current glibc implementation of lock elision using transactional memory instructions.

Uploaded by

John evans
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
56 views

Lock Elision Pthreads and MySQL

The document discusses lock elision, which is a technique supported in recent versions of glibc that allows certain lock acquisitions to be performed without actually acquiring the lock in cases where no contention is detected. The author is curious if this technique provides benefits for databases like MySQL and RocksDB that can experience contention on read-write locks when reads are frequent. They also note that documentation for glibc threading features like this can be limited, and provide a helpful summary from an expert on the current glibc implementation of lock elision using transactional memory instructions.

Uploaded by

John evans
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

smalldatum.blogspot.

com

Lock elision, pthreads and MySQL

3-4 minutes

Yesterday I learned that

lock elision
is supported in

recent versions of glibc


for pthread mutex and rw-lock. I am curious if anyone has results for
MySQL with it. My memory is that InnoDB can suffer from contention
on a rw-lock, but that is a custom rw-lock not the one included with
glibc. But code above the storage engine uses mutex and maybe rw-
lock from glibc.

A rw-lock where reads dominate can suffer from contention because it


has at least twice the memory writes per lock/unlock pair compared to
a mutex. So when the lock hold time is short a mutex wins even when
exclusive access isn't required. This can often be seen in PMP output
where there are convoys and the worst-case is when a thread gets
stuck trying to get the internal latch during unlock, but the InnoDB
custom rw-lock might not have that problem. Lock elision for the rw-
lock might be a big deal in this case.

RocksDB might also benefit from this change.

One of the challenges with glibc pthreads is documentation. I


previously wrote about the difficulty of finding documentation for
PTHREAD_MUTEX_ADAPTIVE_NP. The problem continues. There isn't
much about pthreads in a recent version of the glibc manual. From
Google searches I wasn't able to find recent docs elsewhere, except for
man pages. But man pages don't document
PTHREAD_MUTEX_ADAPTIVE_NP.  With lock elision we get new
options -- PTHREAD_MUTEX_ELISION_NP and
PTHREAD_MUTEX_NO_ELISION_MP. Google searches will take you to
bits of source code and email list discussions. I hope this can be
improved. Given the lack of docs you might need to read the source. I
hope that the community (web-scale companies) can sponsor a tech
writer to provide the missing docs.

There has been drama because the introduction of this feature failed
when it encountered buggy microcode on certain CPUs. Then there
was more drama when it broke buggy software that worked despite the
bugs, until lock elision made the bugs serious. Google searches find
many of the stories.

One of my favorite perks at work is getting answers from experts. In


this case the expert is Nathan Bronson (thank you). A summary of the
glibc 2.23 implementation per the expert is:

NPTL lock elision is performed using TSX's RTM (Restricted


Transactional Memory) instructions XBEGIN, XEND, and XABORT,
rather than TSX's HLE (Hardware Lock Elision) instructions
XACQUIRE and XRELEASE
On x86, elision support is always present when detected by
HAS_CPU_FEATURE(RTM)
pthread_rwlock_t always attempts elision if the hardware has it
(both for .._rdlock and .._wrlock)
pthread_rwlock_t uses an adaptive strategy for falling back to the
non-TSX implementation. If the lock is held in a non-TSX mode,
there is a transaction conflict, or the transaction exceeds TSX's
(undocumented) capacity, then the current lock acquisition and the
3 following use the non-TXN code path. This means that once a
lock falls off the elision path it needs several uncontended
acquisitions before a transaction it will be attempted again. This
seems quite conservative
pthread_rwlock_rdlock -> pthread_rwlock_unlock with a successful
transaction is about twice as fast as the non-TSX implementation
under no contention, and massively better under contention

You might also like