|This article relies on references to primary sources. (December 2011)|
GT.M is a high-throughput key-value database engine optimized for transaction processing. (It is a type also referred to as "schema-less", "schema-free," or "NoSQL.") GT.M is also an application development platform and a compiler for the ISO standard M language, also known as MUMPS.
GT.M, an abbreviation for Greystone Technology M, was developed by the Greystone Technology Corp in the 1980s. It is an implementation of ANSI standard M for various UNIX systems and OpenVMS. In addition to preserving the traditional features of M, GT.M also offers an optimizing compiler that produces object code that does not require internal interpreters during execution.
The database engine, made open source in 2000, is maintained by Fidelity Information Services. GT.M is used as the backend of their FIS Profile banking application, and it powers ING DIRECT banks in the United States, Canada, Spain, France, Italy, Holland and the UK. It is also used as an open source backend for the Electronic Health Record system WorldVistA and other open source EHRs such as Medsphere's OpenVista. It is listed as an open source healthcare solution partner of Red Hat. Today it consists of approximately 2 million lines of code.
GT.M consists of a language subsystem, a database subsystem, and utility programs. The language subsystem and database subsystem are closely integrated, but each is usable without the other. The language and database subsystems share common data organization and typing.
Key-value database files into the TB range (unlimited aggregate database sizes) ACID (Atomic, Consistent, Isolated, Durable) transactions Large scale replication for business continuity Thousands of concurrent users at largest production sites Plug-in architecture for database encryption
Data Organization and Typing
GT.M has only two data types - canonical numbers and strings. A string is any arbitrary sequence of bytes (including nulls). A string such as
"42" is a canonical number. Data typing is dynamic and conversion between the two types is performed on the fly as needed:
1+"42" yields the result
43, and the first character of
There is only one data structure - multi-dimensional sparse arrays (key-value nodes, sub-trees, and associative memory are all equally valid descriptions) with up to 32 subscripts. A scalar can be thought of as an array element with zero subscripts. Nodes with varying numbers of subscripts (including one node with no subscripts) can freely co-exist in the same array. For example, if one wanted to represent the national capitals of the United States:
Set Capital("United States")="Washington"
Set Capital("United States",1774,1776)="Philadelphia"
Set Capital("United States",1776,1777)="Baltimore"
Variables are created on demand when first assigned to. Thus, the first Set command above would create the variable
Capital. Variables have scope in the language, and are called local variables. A database access looks like an array access, for example:
Set ^Capital("United States")="Washington"
but the caret (^) means that it is a database access. Variables used for database access have a single global scope, and of course persist and shared between processes. They are called global variables. The first 31 characters of a variable name are significant.
The Kill and ZKill commands are used to delete subtrees of values.
Numbers in GT.M are accurate to 18 significant digits, of which accuracy to 3 digits to the right of the decimal point are assured when there are 15 or fewer digits to the left. Scientific notation is supported for larger numbers.
The logical database of a GT.M process consists of one or more global variable name spaces, each consisting of unlimited number of global variables. For each global variable name space, a global directory maps global variables to the database files where they actually reside. An unlimited number of global variables can fit within one database file; a global variable must fit in one database file.
A database file consists of up to 224M (276,168,704) database blocks. A database block is a multiple of 512 bytes, with a maximum size of 65,024 bytes. Commonly used block sizes are 4KB, 8KB and 16KB - so, with an 8KB block size, an individual global variable can grow to 1,792GB. A global variable node (global variable, subscripts plus value) must fit in one database block and each block has a 16 byte overhead. So, the largest node that will fit in a database with a 4KB block size is 4,080 bytes. A key (global variable plus subscripts) can be up to 255 bytes.
The database engine is daemonless and processes accessing the database operate with normal user and group ids - a process has access to a database file if and only if the ownership and permissions of that database file (plus any layered access control such as SELinux permits access). Each process has within its address space all the logic needed to manage the database, and processes cooperate with one another to manage database files. When a database file is journaled, updates are written to journal files before being written to database files, and in the event of a system crash, database files can be recovered from journal files.
The database engine also supports transaction processing. So, code such as:
A plug-in architecture allows the database to be encrypted in order to protect data at rest. GT.M is distributed with reference plug-in that uses GnuPG.
Unlike the database where global variable nodes must fit within a database block, local variable strings can grow to 1MB. The GT.M run-time provides dynamic storage allocation with garbage collection. The number of local variables and the number of nodes in local variables are limited only by storage available to the process. The default scope of a local variable is the lifetime of a process. Local variables created within routines using the New command have more limited scope.
GT.M routines are dynamically compiled and linked for execution in the address space of each process. With the exception of the 32-bit implementation of GT.M for the x86 GNU/Linux platform, object modules can also be placed in shared libraries with the standard
ld command, in which case the memory used is shared. This is important because an application such as VistA has over 20,000 routines whose compiled object code exceeds 200MB. A large hospital running VistA can have thousands of concurrently running user processes.
In GT.M, M code can freely call out to C code (or code in other languages with a C compatible interface), and C code can freely call in to M code (so the top level program can be a C
main()). For example is a GT.M module in CPAN, m_python for access from Python or EGTM binding for Erlang.
GT.M is fully supported on the following platforms (in alphabetic order):
- AIX on IBM System p
- GNU/Linux on IBM System z, Itanium, x86_64 and IA-32 (x86) architectures
- HP-UX on Itanium
- Solaris on SPARC
- z/OS on IBM System z
GT.M is also supported on the following platforms. Although bugs are fixed, releases get new functionality only when the code changes are portable to the platforms with no extra work:
There also exists a set of experimental modifications of Linux code base aimed to bring GT.M onto FreeBSD/x64, DragonFlyBSD/x64 and OpenIndiana/x64 (a fork of Solaris for x86/x64), mentioned in order of porting status from the best down to the worst. The code for all the mentioned platforms is to be found on the IDEA labs website.
On GNU/Linux on x86-64 & IA-32 (x86), and on OpenVMS on Alpha/AXP, GT.M is released as Free / Open Source Software (FOSS) under the terms of the GNU Affero General Public License, version 3. On other platforms, it is available under proprietary licenses.
GT.M is predominantly used in healthcare and financial services industry. The first production use of GT.M was in 1986 at the Elvis Presley Memorial Trauma Center in Memphis, Tennessee. Through FIS Profile, it powers ING DIRECT banks in the United States, Canada, Spain, France and Italy.
- Ignacio Valdes (November 17, 2002), K.S. Bhaskar Receives 2002 LMN Achievement Award, linuxmednews.com