Quickdocs

Q

stmx

2024-10-12

Composable Transactional Memory

Upstream URL

github.com/cosmos72/stmx

Author

Massimiliano Ghilardi

License

LLGPL
README
Provided Systems

STMX

Summary

STMX is a high-performance implementation of composable Transactional Memory (TM) for Common Lisp. TM is a concurrency control mechanism aimed at making concurrent programming easier to write and understand. Instead of traditional lock-based programming, one programs with atomic memory transactions, which can be composed together to make larger atomic memory transactions.

A memory transaction gets committed if it returns normally, while it gets rolled back if it signals an error (and the error is propagated to the caller).

Finally, memory transactions can safely run in parallel in different threads, are re-executed from the beginning in case of conflicts or if consistent reads cannot be guaranteed, and their effects are not visible from other threads until they commit.

Memory transactions give freedom from deadlocks, are immune to thread-safety bugs and race conditions, provide automatic roll-back on failure, and aim at resolving the tension between granularity and concurrency.

News

Latest news, 1st March 2020

Fixed STMX for internal changes in SBCL 2.0.0 and SBCL 2.0.2. Updated list of Intel CPUs supporting memory transactions in hardware (Intel TSX) - see below.

News, 16th January 2015

Version 2.0.1 released. It adds support for transactional structs in addition to transactional CLOS objects, and a faster, struct-based implementation of transactional CONS cells and lists, including several list-manipulating functions - see util/tcons.lisp and util/tlist.lisp

Unluckily, the hardware bug that prompted Intel to disable hardware transactional memory (TSX) in August 2014 is still there, and very few new models are available without the bug. So for the moment STMX will be software-only on many CPUs.

Older news

See the file NEWS.md

General documentation

An introduction is available to explain more in detail what STMX is, what it is not, and how it is implemented.

For background information, Composable Memory Transactions is a very good - though a bit technical - explanation of memory transactions and how they are used and combined. For the interested reader, it also goes in deep detail on how to actually implement them.

Supported systems

STMX is currently tested on the following Common Lisp implementations:

  • SBCL

    • version 2.0.9 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 2.0.6 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 2.0.0 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.5.4 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.4.16 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.3.19 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.2.14 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.1.15 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.2.6 (x86) on Debian GNU/Linux stretch (x86_64)
    • version 1.2.8 (armhf) on Debian GNU/Linux wheezy (armhf) inside Qemu
    • version 1.1.15 (powerpc) on Debian GNU/Linux stretch (powerpc) inside Qemu
    • version 1.2.7 (x86_64) on Windows 7 (x86_64)

    Versions < 1.2 have too old builtin ASDF, and must be manually upgraded to ASDF >= 3.1 to load STMX.

  • ABCL

    • version 1.7.1 with OpenJDK 11.0.8 (x86_64) on Debian GNU/Linux bullseye (x86_64)

  • CCL

    • version 1.12 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.11 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.10 (x86_64) on Debian GNU/Linux bullseye (x86_64)
    • version 1.12 (x86) on Debian GNU/Linux bullseye (x86_64)
    • version 1.11 (x86) on Debian GNU/Linux bullseye (x86_64)
    • version 1.10 (x86) on Debian GNU/Linux bullseye (x86_64)
    • version 1.10 (linuxarm) on Debian GNU/Linux wheezy (armhf) inside Qemu
    • version 1.9-r15761 (linuxppc) on Debian GNU/Linux wheezy (powerpc) inside Qemu

  • CLISP

    • version 2.49.92 (x86_64) on Debian GNU/Linux bullseye (x86_64)

    CLISP lacks multi-threading, and its builtin ASDF must be manually upgraded to ASDF >= 3.1 to load STMX.

Partially supported systems

  • CMUCL

    • versions <= 20e should be started with the command line options "-fpu" "x87"

    to improve STMX reliability, see doc/supported-systems.md.

    Version >= 20f no longer support the command line options "-fpu" "x87" and STMX runs reliably only in single-thread mode.

    CMUCL has too old builtin ASDF, and must be manually upgraded to ASDF >= 3.1 to load STMX.

  • ECL

    • version 16.1.2 (x86_64) on Debian GNU/Linux stretch (x86_64)
    • version 16.0.0 (x86_64) on Debian GNU/Linux stretch (x86_64)
    • version 15.2.21 (x86_64) on Debian GNU/Linux stretch (x86_64)
    • version 13.5.1 (x86_64) on Debian GNU/Linux stretch (x86_64)

    There are known issues running STMX on ECL, see Github issues

    ECL versions 16.0.0 and 16.1.2 run STMX reliably enough for lightweight use, but sometimes still hang in test suite.

Untested systems

STMX will probably work on several other Common Lisp implementations as long as they support log4cl, closer-mop, bordeaux-threads and trivial-garbage, but the author gives no guarantees.

Installation and loading

Stable version - from Quicklisp

STMX is available from Quicklisp. The simplest way to obtain it is to first install Quicklisp then run these commands from REPL:

CL-USER> (ql:quickload "stmx")
;; lots of output...
CL-USER> (use-package :stmx)

If all goes well, this will automatically download and install the stable branch of STMX and its dependencies:

  • log4cl
  • closer-mop
  • bordeaux-threads
  • trivial-garbage

Latest version - from GitHub

In case you want to use the "latest and greatest" version directly from the author, in order to get the newest features - most notably hardware memory transactions - improvements, bug fixes, and occasionally new bugs, you need to download it into your Quicklisp local-projects folder. Open a shell and run the commands:

$ cd ~/quicklisp/local-projects
$ git clone git://github.com/cosmos72/stmx.git

then proceed as before - load a REPL and run:

CL-USER> (ql:quickload "stmx")
;; lots of output...
CL-USER> (use-package :stmx)

If all goes well, it will automatically load STMX and its dependencies.

Note: unless you know what you are doing, do not try to load different STMX versions one after the other from the same REPL - strange things may happen.

Other versions - from Sourceforge

All the stable versions of STMX, present and past, are also available from Sourceforge, including version 1.9.0.

Troubleshooting

In case you get errors:

  • check that Quicklisp is installed correctly, for example by executing at REPL:

      CL-USER> (ql:quickload "closer-mop")

  • if you tried to download the stable version from Quicklisp, check that your quicklisp is updated and knows about STMX:

      CL-USER> (ql:system-apropos "stmx")

    should print something like

      #<SYSTEM stmx / stmx-stable-7e68763b-git / quicklisp 2013-06-15>
  #<SYSTEM stmx.test / stmx-stable-7e68763b-git / quicklisp 2013-06-15>

    If it doesn't, you need to update Quicklisp as described here - search for "To get updated software" in the page.

  • if you tried to download the latest version from GIT, check that you downloaded STMX creating an stmx/ folder inside your Quicklisp local-projects folder, usually ~/quicklisp/local-projects

Testing that it works

After loading STMX for the first time, it is recommended to run the test suite to check that everything works as expected. From the REPL, run:

CL-USER> (ql:quickload "stmx.test")
;; lots of output...
CL-USER> (fiveam:run! 'stmx.test:suite)
;; even more output...
 Did 7133 checks.
    Pass: 7133 (100%)
    Skip: 0 ( 0%)
    Fail: 0 ( 0%)

Note: (ql:quickload "stmx.test") intentionally works only after (ql:quickload "stmx") has completed successfuly.

The test suite should report zero Skip and zero Fail; the number of Pass may vary. You are welcome to report any failure you get while running the test suite, please include in the report:

  • operating system name and version (example: Debian GNU/Linux x86_64 version 7.0)
  • Common Lisp implementation and version (example: SBCL 1.0.57.0.debian, x86_64)
  • EXACT output produced by the test suite
  • any other relevant information

See "Contacts, help, discussion" below for the preferred method to send the report.

Basic usage

STMX offers the following Lisp macros and functions, also heavily documented in the sources - remember (describe 'some-symbol) at REPL.

  • TRANSACTIONAL declares that a class or struct is transactional, i.e. that its slots contain transactional data. Use it to wrap a class or a struct definition:

      (transactional
    (defclass foo ()
      ((value1 :type integer :initarg :value1 :initform 0)
       (value2 :type string  :initarg :value2 :initform ""))))

  (transactional
    (defstruct bar ()
      (value1 0  :type integer)
      (value2 "" :type string)))

    If you want to declare a slot as non-transactional, for example because it is immutable, add the option :transactional nil:

      (transactional
    (defclass tred-black-tree ()
      ((root           :type t)
       (key-comparator :type function :transactional nil))))

    Note: on all Common Lisp implementations listed above, CLOS slot accessors have been tested to work correctly, i.e. they honour the transactional machinery (implemented with MOP slot-value-using-class) and return the same values as SLOT-VALUE.

    In the past, some cases were found - and fixed - where slot accessors would not work correctly on transactional classes. It is strongly recommended to run the test suite when using STMX for the first time on a system: it will also check that accessors work on transactional classes.

    Support for (TRANSACTIONAL (DEFSTRUCT ...)) was added in STMX version 2.0.1 on January 2015. Previously, only (TRANSACTIONAL (DEFCLASS ...)) was supported.

  • ATOMIC is the main macro: it wraps Lisp forms into an atomic memory transaction then executes them. For example, defining

      (defun show-foo (obj)
    (declare (type foo obj))
    (multiple-value-bind (value1 value2)
        (atomic
          (values (slot-value obj 'value1)
                  (slot-value obj 'value2)))
      (format t "atomic function show-foo: foo contains ~S, ~S~%"
              value1 value2)))

  (defmethod set-foo ((obj foo) value1 value2)
    (declare (type integer value1)
             (type string value2))
    (atomic
      (setf (slot-value obj 'value1) value1)
      (setf (slot-value obj 'value2) value2))
    (format t "atomic method set-foo: foo now contains ~S, S~%"
            value1 value2))

    SHOW-FOO will atomically read the slots VALUE1 and VALUE2 of a FOO instance, then print both. Note that (format t ...) is outside the atomic block - more on this later.

    SET-FOO will atomically set the slots VALUE1 and VALUE2 of a FOO instance.

    Using these two functions, STMX guarantees that multiple threads accessing the same FOO instance will always see consistent values for both slots, i.e. SHOW-FOO will never see intermediate states of a transaction, where for example one slot has been updated by SET-FOO, but the other slot has not been updated yet.

    This is the main feature of STMX: if an atomic block completes normally, it is assumed to be successful and it gets committed: all its writes to transactional memory become visible simultaneously to other threads. If instead an atomic block exits with a non-local control transfer (signals an error, throws, or invokes a (go some-label)), it is assumed to be failed and it gets rolled back: all its writes to transactional memory are discarded.

    Warning: in order to avoid deadlocks and conflicts while still reaching good performance, STMX may execute more than once the contents of an atomic block. Also, some instructions as (retry) described below, explicitly cause an atomic block to be re-executed from the beginning. For this reasons, atomic blocks should not contain irreversible operations such as input/output. More details in the paragraph INPUT/OUTPUT DURING TRANSACTIONS below.

    Note: new threads must be created with (bordeaux-threads:make-thread) in order to establish thread-local bindings needed by STMX. A safety check that detects missing thread-local bindings has been recently added to STMX.

    Note: STMX allows using transactional data both inside and outside atomic blocks, but be aware that accessing transactional data from outside atomic transactions is only intended for debugging purposes at the REPL: in a program it can cause a lot of problems, due to inconsistencies and due to other threads not being notified when a transactional memory location is updated. Future versions may remove this convenience hack and replace it with a cleaner, stricter mechanism. In a program, always make sure that all code that accesses transactional data is directly or indirectly executed inside an (atomic ...) block.

  • TRANSACTION declares that a method or function is an atomic memory transaction, and is actually just a macro that wraps the body of a function or method in an (atomic ...) block. In the past, it was suggested as a more convenient alternative to ATOMIC, but for various stylistic reasons the current recommendation is to avoid it. The main reason is that it encourages performing too many operations inside an atomic block, including irreversible ones as input/output, which has impredictable behaviour and should be really avoided. Examples:

      (transaction
    (defun get-foo-values (obj)
      (declare (type foo obj))
      (values
        (value1-of obj) (value2-of obj))))

  (transaction
    (defmethod set-foo-values ((obj foo) value1 value2)
      (declare (type integer value1)
               (type string value2))
      (setf (value1-of obj) value1)
      (setf (value2-of obj) value2)
      obj))

  • Composing transactions

    A key feature of ATOMIC is its composability: smaller transactions can be composed to create larger transactions. For example, the following three program fragments are perfectly equivalent:

    1. use (atomic ...) to wrap into a single transaction many smaller (atomic ...) blocks

      (defmethod swap-value1-of ((x foo) (y foo)) (format t "swapping value1 of ~S and S%" x y) (atomic (rotatef (slot-value x 'value1) (slot-value y 'value1))))

      (defmethod swap-value2-of ((x foo) (y foo)) (format t "swapping value2 of ~S and S%" x y) (atomic (rotatef (slot-value x 'value2) (slot-value y 'value2))))

      (defmethod swap-contents ((x foo) (y foo)) (atomic (swap-value1-of x y) (swap-value2-of x y)))

    2. write redundant (atomic ...) blocks

      (defmethod swap-contents ((x foo) (y foo)) (format t "swapping value1 and value2 of ~S and S%" x y) (atomic (atomic (rotatef (slot-value x 'value1) (slot-value y 'value1))) (atomic (rotatef (slot-value x 'value2) (slot-value y 'value2)))))

    3. write a single (atomic ...) block

      (defmethod swap-contents ((x foo) (y foo)) (format t "swapping value1 and value2 of ~S and S%" x y) (atomic (rotatef (slot-value x 'value1) (slot-value y 'value1)) (rotatef (slot-value x 'value2) (slot-value y 'value2))))

    This composability property has an important consequence: transactional code, possibly written by different people for unrelated purposes, can be combined into larger transactions without modifying it - actually, without looking at the source code at all - as long as it all uses the same transactional library.

    The STMX machinery will guarantee that transactions intermediate status, where an atomic block is half-way through its job, will never be visible to other transactions.

    For example, it becomes trivial to write some code that atomically removes an object from a transactional container and adds it to another one: just write something like

      (defmethod move-obj-from-a-to-b ((a some-container) (b another-container))
    (atomic
      (let ((obj (take-obj-from-some-container a)))
         (put-obj-into-another-container obj b))))

    and it will work as long as both container are transactional and use the same transaction library (in this case, STMX).

    A lot of facts that in other concurrent programming paradigms can be great obstacles to such a solution become completely irrelevant when using transactions: it is irrelevant that the two containers may be unrelated classes, that the respective authors may not have anticipated such need in the APIs, that the internal details of the two implementations may be unknown to the author of code that combines them atomically (the move-obj-from-a-to-b in the example), that other existing code in the program uses the same containers a and b but does not cooperate with move-obj-from-a-to-b.

    Style suggestion: in order to guarantee that all transactional memory accesses are performed inside an atomic block, it may be tempting to wrap each function or method body inside (atomic ...). While safe and correct, this approach has a small performance penalty that performance-critical code may want to avoid by minimizing the number of (atomic ...) blocks: it is enough to have a top-level atomic block that corresponds to the largest transaction that one wants to execute, and omit inner atomic blocks in the same or other functions called directly or indirectly from the top-level atomic block. In such case, it is strongly recommended to insert in the documentation of the functions accessing transactional memory without a direct atomic block a sentence like "This function should be always invoked from inside an STMX atomic block."

  • RETRY is a function. It is more tricky to understand, but really powerful. As described in the summary, transactions will commit if they return normally, while they will rollback if they signal an error or condition.

    The (retry) function call offers a third option: if invoked inside a transaction, it tells STMX that the transaction cannot complete immediately, for example because some necessary data is not currently available, and instructs STMX to wait until the data has changed, then re-execute the transaction from scratch.

    How does (retry) know which data it should monitor for changes? Simple: it will monitor all transactional data (including slots of transactional objects) that was read since the beginning of the transaction and until (retry) was invoked.

    With RETRY, reliable communication among threads is (hopefully) extremely simple to implement: a thread can read one (or more) transactional data, checking for values that some other thread will write there, and just (retry) if no appropriate values are there yet.

  • ORELSE is a macro to execute two or more Lisp forms as alternatives in separate, nested transactions: if the first retries or detects an inconsistent read, the second will be executed and so on, until one transaction either commits (returns normally) or rollbacks (signals an error or condition). It can only be used inside a transaction.

  • NONBLOCKING is an utility macro based on ORELSE to convert a blocking transaction into another that returns NIL instead of waiting (and otherwise returns T followed by the values or the original transaction)

      (nonblocking (x) (y) (z))

    basically expands to

      (orelse (values t (progn (x) (y) (z))) nil)

    with the difference that (nonblocking ...) actually captures all the values returned by the transaction, not just the first as in the example above.

Input/Output during transactions

WARNING: since transactions will be re-executed in case of conflicts with others and can also rollback or retry, all code inside an atomic block may be executed more times than expected, or may be executed when not expected.

Some transactional memory implementations, especially for statically-typed languages, forbid performing input/output during a transaction on the ground that I/O is not transactional: if a transaction sends an irreversible command to the outside world, there is no way to undo it in case the transaction rolls back, retries or conflicts.

STMX does not implement such restrictions, i.e. I/O and any other irreversible action can also be performed inside an atomic block. This means you are free to launch missiles during a transaction, and destroy the world when you shouldn't have. You have been warned.

Despite the risk, there are at least two reasons for such a design choice:

  • Forbidding I/O operations inside transactions, if done at all, should be done while compiling a program rather than while running it. In Common Lisp, neither of the two seems easy to implement.
  • Common Lisp programs are often much more dynamic and flexible than programs in other languages, and programmers are trusted to know what they are doing. Such a prohibition does not seem to fit well with this spirit.

The typical solution for the above risk is: during a transaction, perform I/O only for debugging purposes, for example using a logging library as log4cl (or whatever is appropriate for your program), and queue any I/O operation in a transactional buffer. Then, invoke a separate function that first runs a transaction to atomically consume the buffer and only later, outside any transaction, performs the actual I/O operation.

An alternative solution is: during a transaction, instead of performing I/O pass to AFTER-COMMIT a function that will perform I/O when executed. Note: AFTER-COMMIT is described in Advanced usage below, read it carefully because functions executed by AFTER-COMMIT have several restrictions on what they are allowed to do.

Advanced usage

For those cases where the basic features are not sufficient, or where more control is desired during the execution of transactional code, some advanced features are available:

  • RUN-ATOMIC is the function version of ATOMIC: takes a single function argument and executes it in a transaction. This means the following two code snippets are equivalent:

      (defvar a (make-instance 'foo))
  (defvar b (make-instance 'foo))
  (atomic
    (set-foo a 1 "abc")
    (set-foo b 2 "def"))

    and

      (defvar a (make-instance 'foo))
  (defvar b (make-instance 'foo))

  (defun init-foo-a-and-b ()
    (set-foo a 1 "abc")
    (set-foo b 2 "def"))

  (run-atomic #'init-foo-a-and-b)

  • RUN-ORELSE is the function version of ORELSE: it accepts any number of functions and executes them as alternatives in separate, nested transactions: if the first retries or is invalid, the second will be executed and so on, until one function either commits (returns normally) or rollbacks (signals an error or condition).

    If X, Y and Z are no-argument functions, the following two lines are equivalent:

      (orelse (x) (y) (z))
  (run-orelse #'x #'y #'z)

  • BEFORE-COMMIT is a macro that registers Lisp forms to be executed later, just before the transaction tries to commit. It can be useful to normalize or simplify some transactional data, or perform any kind of bookkeeping activity.

    Be aware that the transaction is not yet committed when the forms registered with BEFORE-COMMIT run. This means in particular:

    • There is no guarantee that the commit will succeed.

    • If the forms signal an error when executed, the error is propagated to the caller, forms registered later with BEFORE-COMMIT are not executed, and the transaction rolls back.

    • The forms can read and write normally to transactional memory, and in case of conflicts the whole transaction, including all forms registered with BEFORE-COMMIT, is re-executed from the beginning.

    • The forms cannot (retry) - attempts to do so will signal an error. Starting a nested transaction and retrying inside that is acceptable, as long as the (retry) does not propagate outside the forms themselves.

  • AFTER-COMMIT is another macro that registers Lisp forms to be executed later, but in this case they are executed immediately after the transaction has been successfully committed. It can be useful to notify some subsystem that for any reason cannot call (retry) to be informed of changes in transactional memory - for example because it is some existing code that one does not wish to modify.

    In this case, the transaction is already committed when the forms registered with AFTER-COMMIT run, and (since STMX 1.3.2) the forms are executed outside any transaction. There are some limitations on what the forms can do:

    • If the forms signal an error when executed, the error is propagated to the caller, forms registered later with AFTER-COMMIT are not executed, but the transaction remains committed.

    • The forms are not executed inside a transaction: while it is certainly possible to explicitly run an (atomic) block from them, doing so would probably defeat the purpose of AFTER-COMMIT and it may also cause a significant performance penalty.

  • CALL-BEFORE-COMMIT is the function version of BEFORE-COMMIT: it accepts a single function and registers it to be executed before the transaction tries to commit.

  • CALL-AFTER-COMMIT is the function version of AFTER-COMMIT: it accepts a single function and registers it to be executed after the transaction has been successfully committed.

  • TVAR is the class implementing transactional memory behind the scenes. It is used internally by slots of transactional classes, but can also be used directly. Except if specified, all its functions and methods work both inside and outside transactions (remember that using transactional memory outside transactions is only intended for debugging purposes). Functions and methods:

    • (tvar [initial-value]) Create a new TVAR, optionally bound to a value.
    • ($-slot var) Get the value of VAR. Signals an error if VAR is not bound to any value. Note: before STMX 1.9.0, this function was named ($ var).
    • (setf ($-slot var) value) Store VALUE into VAR. Note: before STMX 1.9.0, this function was named (setf ($ var) value).
    • (bound-$? var) Return true if VAR is bound to some value.
    • (unbind-$ var) Unbind VAR from its value.
    • (value-of var) getter method, equivalent to ($-slot var)
    • (setf (value-of var) value) setter method, equivalent to (setf ($-slot var) value)

    For programmers that want to squeeze the last CPU cycle out of STMX, there are also some more specialized functions:

    • ($ var) Get the value of VAR. Return +unbound-tvar+ if VAR is not bound to any value.
    • (setf ($ var) value) Set the value of VAR. Identical to (setf ($-slot var) value) and provided for simmetry with ($ var).

Hardware transactions

STMX versions 1.9.0 or later can take advantage of hardware transactions on Intel CPUs that support Transactional Synchronization Extensions (TSX) As of February 2020, many recent consumer and server Intel CPUs support them, including at least:

  • 7th generation Core i5: 7500 7500T 7600 7600K 7600T 7Y57

  • 8th generation Core i5: 8500 8500T 8600 8600T 8600K

  • 9th generation Core i5: 9500 9500E 9500F 9500T 9500TE 9600 9600K 9600KF 9600T

  • 10th generation Core i7: -

  • 7th generation Core i7: 7600U 7660U 7700 7700K 7700T 7820EQ 7820HK 7820HQ 7920HQ 7Y75

  • 8th generation Core i7: 8086K 8650U 8665U 8665UE +8700 8700 8700B 8700K 8700T 8706G 8850H

  • 9th generation Core i7: 9700 9700E 9700F 9700K 9700KF 9700T 9700TE 9850H 9850HE 9850HL

  • 10th generation Core i7: -

  • 8th generation Core i9: 8950HK

  • 9th generation Core i9: 9880H 9900 9900K 9900KF 9900KS 9900T 9980HK

(This list is necessarily incomplete. To check whether a specific Intel CPU supports Transactional Synchronization Extensions (TSX) browse https://ark.intel.com/)

To actually use hardware transactions from STMX, there are two more requirements:

  • a 64-bit version of SBCL - any version >= 1.2.0 should work
  • a 64-bit unix-like operating system - at the moment only Linux x86_64 is tested

Also, hardware transactions only work in compiled code - SBCL sometimes interprets very short functions and simple code executed at REPL instead of compiling them, which may cause hardware transactions to fail.

How to tell if hardware transactions are supported

There are several ways. The easiest are:

  • From outside transactions, run the macro (HW-TRANSACTION-SUPPORTED?). It internally calls the CPUID assembler instruction and returns T if hardware transactions are supported, or NIL if they are not.
  • Try to use them, for example by executing (ATOMIC (HW-TRANSACTION-SUPPORTED-AND-RUNNING?)) in compiled code - hardware transactions typically do not work in interpreted code. Thus actually execute something like (DEFUN HW-TRANSACTION-TEST () (ATOMIC (HW-TRANSACTION-SUPPORTED-AND-RUNNING?))) (HW-TRANSACTION-TEST)

How to use hardware transactions

STMX automatically uses hardware transactions if they are supported. There is no need for special commands, just execute the usual (ATOMIC ...) or (RUN-ATOMIC ...) forms.

Hardware transactions have several limitations, and STMX will seamlessly switch to (slower) software transactions in the following cases:

  • hardware limits are exceeded, for example read-set or write-set are larger than CPU L1 cache

  • executing a function or macro not supported by hardware transactions. The list is subject to change, it currently includes:

    • STMX functions and macros: RETRY, ORELSE, RUN-ORELSE, BEFORE-COMMIT, AFTER-COMMIT, CALL-BEFORE-COMMIT, CALL-AFTER-COMMIT
    • any Common Lisp function or macro that signals an error, or allocates non-trivial amounts of memory, or performs any kind of system calls, including input/output, sleeping and context switching.

  • executing a CPU instruction not allowed inside hardware transaction. In particular, Intel TSX guarantees that CPU instructions

    • CPUID, PAUSE, XABORT

    will always abort a hardware transaction, but many other CPU instructions typically have the same effect, including possibly:

    • Calls to the operating system and returns from it: SYSENTER, SYSCALL, SYSEXIT, SYSRET.
    • Interrupts: INT n, INTO.
    • Input/Output: IN, INS, REP INS, OUT, OUTS, REP OUTS and their variants.
    • All X87 and MMX instructions. On the opposite, XMM and YMM registers and the MXCSR register can be used inside a hardware transaction.
    • CLI, STI, POPFD, POPFQ, CLTS.
    • Instructions that update segment registers, debug registers and/or control registers such as DF (CLD and STD instructions), DS/ES/FS/GS/SS and CR0/CR2/CR3/CR4/CR8.
    • TLB and Cacheability control: CLFLUSH, INVD, WBINVD, INVLPG, INVPCID, and memory instructions with a non-temporal hint (MOVNTDQA, MOVNTDQ, MOVNTI, MOVNTPD, MOVNTPS, and MOVNTQ).
    • Processor state save: XSAVE, XSAVEOPT, and XRSTOR.
    • VMX instructions: VMPTRLD, VMPTRST, VMCLEAR, VMREAD, VMWRITE, VMCALL, VMLAUNCH, VMRESUME, VMXOFF, VMXON, INVEPT, and INVVPID.
    • SMX instructions: GETSEC.
    • Miscellaneous: UD2, RSM, RDMSR, WRMSR, HLT, MONITOR, MWAIT, XSETBV, VZEROUPPER, MASKMOVQ, and V/MASKMOVDQU.

    For details and up-to-date information, see Intel Instruction Set Programming Reference, Chapter "Transactional Synchronization Extensions".

Utilities and examples

See the example and util folder, which contains several examples and utilities built with STMX and should be relatively straightforward to understand. The folder util contains the following classes with related methods and functions, all in the STMX.UTIL package - for more details, use (describe 'some-symbol) at REPL:

  • TCELL is the simplest transactional class. It is created with (tcell [initial-value]) and it can be empty or hold a single value.

    Methods: FULL? EMPTY? EMPTY! PEEK TAKE PUT TRY-TAKE TRY-PUT.

    When empty, taking a value will (retry) and wait until some other thread puts a value.

    When full, putting a value will (retry) and wait until some other thread removes the current value.

    Note: raw TVARs support exactly the same methods.

  • TCONS is a transactional cons cell. It is created with (tcons first-value second-value).

    Functions: TFIRST (SETF TFIRST) TREST (SETF TREST).

    Seldom used directly.

  • TLIST is a transactional list, composed of TCONS cells. It is created with (tlist [values ...]).

    Functions: TFIRST (SETF TFIRST) TREST (SETF TREST) TPUSH TPOP TSECOND TTHIRD TFOURTH TNTH TLAST TLIST-LENGTH TLIST* and many others. See util/tlist.lisp for details.

    Normal lists are perfectly suitable for transactional use as long as they are not destructively modified, so TLIST is often unnecessary: it becomes needed only to support transactional destructive modifications.

  • TSTACK is a transactional first-in-last-out buffer. It is created with (tstack) and it can be empty or hold unlimited values.

    Methods: FULL? EMPTY? EMPTY! PEEK TAKE PUT TRY-TAKE TRY-PUT.

    All methods append or remove values from the end, and putting a value always succeeds, even when other values are already present: the new value is simple appended at the end. For the rest, the methods behave as described for the TCELL class.

  • TFIFO is a transactional first-in-first-out buffer. It is created with (make-instance 'tfifo) and it can be empty or hold unlimited values.

    Methods: FULL? EMPTY? EMPTY! PEEK TAKE PUT TRY-TAKE TRY-PUT.

    PUT and TRY-PUT append values at the end, PEEK TAKE and TRY-TAKE get or remove them from the beginning, shifting the remaining values. For the rest, the methods behave as described for the TCELL and TSTACK classes.

  • TCHANNEL is a transactional multicast channel. It is created with (make-instance 'tchannel), can contain unlimited values and it is write-only. To read from it, create a TPORT as described below.

    Methods: FULL? EMPTY? PUT TRY-PUT.

    PUT and TRY-PUT append values at the end, making them available to connected ports. FULL? always returns nil, since a channel can contain unlimited values. EMPTY? always returns t, since it is not possible to get values from a channel.

    It is possible to write into the same channel from multiple threads: added elements will be interleaved and made available to all connected ports.

  • TPORT is a transactional reader for TCHANNEL. It is created with (make-instance 'tport :channel some-channel). Ports do not support putting values, they are used to retrieve values from the channel they are connected to.

    Methods: FULL? EMPTY? EMPTY! PEEK TAKE TRY-TAKE.

    PEEK TAKE and TRY-TAKE get or consume values previously added to the connected channel. All ports connected to the same channel receive all the values in the same order, and they consume values independently: taking a value from a port does not consume it from the other ports.

    FULL? always returns t, since it is not possible to put values in a port. EMPTY? returns t if some values are available to read or consume. EMPTY! consumes all values currently available.

    It is also possible to use the same port from multiple threads: elements consumed by one thread will not be available to other threads using the same port.

  • THASH-TABLE is a transactional hash table. It is created with (make-instance 'thash-table [:test 'some-test-function] [:hash 'some-hash-function]).

    One difference from standard Common Lisp HASH-TABLE:

    • a hash function can be specified explicitly with :hash 'some-hash-function

    For the usual test functions, i.e. 'eq 'eql 'equal and 'equalp the hash function can be omitted and a safe default (usually 'sxhash) will be used. For other test functions, the hash function becomes mandatory.

    Methods: GHASH-TABLE-COUNT GHASH-TABLE-COUNT> GHASH-TABLE-COUNT<= GHASH-TABLE-EMPTY? CLEAR-GHASH GET-GHASH (SETF GET-GHASH) SET-GHASH REM-GHASH MAP-GHASH DO-GHASH COPY-GHASH GHASH-KEYS GHASH-VALUES GHASH-PAIRS GHASH-TEST GHASH-HASH.

    Warning: retrieving the number of elements in a transactional container is potentially expensive: to maintain consistency, it inhibits concurrent insertion and removal from other threads. For this reason, use GHASH-TABLE-COUNT sparingly.

    THASH-TABLE constructor arguments test and hash changed in STMX 2.0.0. They now must be function names (i.e. symbols), previously they were actual functions.

  • TMAP is a transactional sorted map, backed by a red-black tree. It is created with (make-instance 'tmap :pred compare-function) where COMPARE-FUNCTION must be the name of a function accepting two arguments, KEY1 and KEY2, and returning t if KEY1 is smaller that KEY2. For numeric keys, typical COMPARE-FUNCTIONs are '< or '> and the faster 'fixnum< or 'fixnum>. For string keys, typical COMPARE-FUNCTIONs are 'string< and 'string>.

    Note: COMPARE-FUNCTIONs changed in STMX 2.0.0. They now must be function names (i.e. symbols), previously they were actual functions.

    Methods: GMAP-PRED GMAP-COUNT GMAP-EMPTY? CLEAR-GMAP GET-GMAP (SETF GET-GMAP) SET-GMAP REM-GMAP MIN-GMAP MAX-GMAP MAP-GMAP DO-GMAP GMAP-KEYS GMAP-VALUES GMAP-PAIRS.

    Warning: retrieving the number of elements in a transactional container is potentially expensive: to maintain consistency, it inhibits concurrent insertion and removal from other threads. For this reason, use GMAP-COUNT sparingly.

  • GHASH-TABLE is the non-transactional version of THASH-TABLE. Not so interesting by itself, as Common Lisp offers a standard (and usually faster) HASH-TABLE implementation. It supports exactly the same methods as THASH-TABLE.

  • RBMAP is the non-transactional version of TMAP. Not so interesting by itself, as many other red-black trees implementations exist already on the net. It supports exactly the same methods as TMAP.

Performance

STMX automatically discovers and takes advantage of many optional, non-standard features of the underlying Common Lisp compiler. It also performs graceful degradation, i.e. if the fastest version of a feature is not available it automatically switches to a slower, available alternative.

Depending on the available features, STMX performance can vary up to a factor 100 or more (!).

To reach its peak performance, several requirements need to be satisfied by the hardware and by the Lisp compiler being used. They are listed here in order of importance:

Hardware requirements:

  • support hardware transactions (Intel TSX). Without them, STMX is at least 4-5 times slower. Or, if you prefer since Intel TSX is currently very rare, with it STMX is at least 4-5 times faster. As of August 2013, STMX can use hardware transactions only on 64-bit SBCL.

Lisp compiler requirements:

  1. it must have good multi-threading support. Without it, what would you need a concurrency library as STMX for?
  2. it must expose atomic compare-and-swap operations, to implement fast mutexes. A much slower alternative, but still better than nothing, is to expose a function that returns which thread has acquired a bordeaux-threads lock.
  3. it must produce fast, highly optimized code.
  4. it must be 64-bit. 32-bit is much slower because transactional memory version counters are then BIGNUMs instead of FIXNUMs.
  5. it must expose memory barrier operations. This is less important on x86 and x86-64, and more important on unordered architectures (almost all others).

Among the non-commercial Lisp compilers, SBCL is the only one known to STMX author that satisfies all the compiler requirements, and (guess why) the only one where STMX author has implemented support for hardware transactions.

Actually, all the other tested free Lisp compilers (ABCL, CCL, CMUCL, ECL) are at least somewhat lacking in the area "fast, highly optimized code", and none of them offers atomic compare-and-swap or memory barrier operations at all. One - CMUCL - produces relatively fast code, but does not support native threads. STMX is not tested on any commercial Lisp compiler, so performance on them is simply unknown.

For these reasons, STMX will reach the highest known performance on SBCL by a large margin - possibly by a factor from 10 to 100 or more with respect to other tested systems.

For more performance considerations and a lot of raw numbers produced by running micro-benchmarks, see the included files doc/benchmark.md, doc/benchmark-abcl.md, doc/benchmark-ccl64.md and doc/benchmark-cmucl.md.

The short version is: as of March 2015, on a fast consumer PC (Core i7 4770 @ 3.5GHz or better) with 64-bit SBCL 1.1.9 or better, STMX can execute more than 35 millions hardware transactions per second per CPU core, and more than 7 millions software transactions per second per CPU core. The second platform in terms of performance is CCL (x86_64), that reaches 1.1 millions software transactions per second per CPU core using two threads, but STMX performance quickly decreases with more threads (reason still needs to be investigated).

A small example with very short transactions is the dining philosophers, with 5 reads and 5 writes to transactional memory per atomic block, where each CPU core runs approximately 4.5 millions software transactions per second - hyperthreading has very limited effects.

Obviously, performance in other usage scenarios will depend on the complexity of the code inside transactions, on the availability of hardware transactions, on the number of reads and writes to transactional memory, and the rate of conflicts and rollbacks.

Note

These result are not absolute performance considerations of the tested Lisp systems. They are simply the outcome of running micro-benchmarks of a particular library optimized for SBCL (see the hardware transactions, atomic compare-and-swap and memory barriers considerations) on several other Lisp systems. Do not try to construct these results as STMX author's opinions on the mentioned Lisp systems.

Lee-STMX

For a less artificial and hopefully more realistic benchmark, the author has ported Lee-TM, a non-trivial benchmark suite for transactional memory developed in 2007 by the University of Manchester (UK). The result is Lee-STMX - as of July 2013, its status is BETA.

Contacts, help, discussion

As long as the traffic is low enough, GitHub Issues can be used to report test suite failures, bugs, suggestions, general discussion etc.

If the traffic becomes high, more appropriate discussion channels will be set-up.

The author will also try to answer support requests, but gives no guarantees.

Status

As of July 2013, STMX is being written by Massimiliano Ghilardi and is considered by the author to be stable.

STMX is a full rewrite of CL-STM, which has been developed by Hoan Ton-That for the Google Summer of Code 2006.

Donations

STMX is a spare-time project. Donations can help the project by recognizing its usefulness and covering expenses.

You can donate with PayPal or credit card.

Legal

STMX is released under the terms of the Lisp Lesser General Public License, known as the LLGPL.

Dependencies (6)

  • alexandria
  • bordeaux-threads
  • closer-mop
  • fiveam
  • log4cl
  • trivial-garbage

Dependents (1)

  • GitHub
  • Quicklisp