Common Tools Interface

Contents

A. Introduction

B. Common Tools Interface synopsis

C. Library structure

D. General CTI frontend interface

D.1. Function listing

E. Applications

E.1. Application launch

E.2. Application launch with barrier

E.3. Application attach

E.4. WLM-specific Extensions

E.5. Application facilities

E.5.1. Function listing

F. File transfer support

F.1. Sessions

F.1.1. Function listing

F.2. Manifests

F.2.1. Function listing

F.3. Object lifetime

G. Tool daemons

H. Backend library

H.1. Function listing

Introduction

This document is organized as follows. Section B provides a synopsis of the Common Tools Interface (CTI) which is an API designed by Cray to facilitate bootstrapping of tool launch alongside applications. Section C provides an overview of the library layout and associated header files. Sections D through G describe functionality available on the tool frontend. Section H describes functionality available on the tool backend.

CTI is available on github at the following link: https://github.com/common-tools-interface/cti

Common Tools Interface synopsis

The Common Tools Interface (CTI) is an infrastructure framework to enable tool developers to launch, interact with, and run tool daemons alongside applications on HPC systems. CTI provides a collection of interfaces and supporting utilities to export services for use by tools. CTI enables vendors to provide workload manager-specific implementations to rapidly enable support for existing tools that use CTI.

CTI abstracts the process of application launch and lifecycle management by providing an API that allows underlying implementations to transparently handle workload manager-specific details. For example, a SLURM system makes use of srun to launch applications, skill to send signals to a running application, sbcast to ship files to compute nodes, and the MPIR standard to harvest placement information about an application. In contrast, a generic implementation uses ssh to launch applications and send signals, scp to ship files, and might lack MPIR support. CTI provides a common API for developers of tools to run these essential HPC job management tasks across different implementations.

From a tool developer's perspective, CTI manages interaction with an application, tool daemon launch, and staging dependencies to compute nodes. A tool may need to do tasks such as starting daemons on compute nodes alongside an application, make various binaries and shared library dependencies available to the tools on the compute node, use automatically generated support files created on a login node, write temporary files to a safe location, determine the PIDs of application processes, etc.

All CTI functionality is available regardless of system configuration. System specific details such as availability of a parallel file system are transparent from the API level. This allows the underlying implementation to make decisions based on the particular system and in turn allow tools to be extensible in a diverse HPC ecosystem.

Library structure

CTI is split into two libraries for use by tool developers: a frontend library for use on the login node, and a backend library for use by a tool daemon. These libraries have associated pkg-config files to facilitate ease of linking against the associated libraries. Each library has its own associated header file. These associations are listed below.

Library name	Header File	pkg-config file
libcommontools_fe.so	common_tools_fe.h	common_tools_fe.pc
libcommontools_be.so	common_tools_be.h	common_tools_be.pc

The frontend library is for use on the login node. It is used for application launch/registration, obtaining information about the application, and bootstrapping tool daemons. It implements the bulk of CTI functionality.

The backend library is for use by tool daemons. It allows developers to determine node local information such as PIDs of application processes, logical PE placement of application ranks, and filesystem locations of staged files. To use the backend library, it must be linked against a tool daemon launched via cti_execToolDaemon by the frontend library. For more information on tool daemon launch see section Tool daemons.

The header file common_tools_shared.h defines the types shared by the frontend and backend library, as well as environment variables.

General CTI frontend interface

Several functions exist for setting configuration with the CTI frontend, querying error information, or other information about the login node that doesn't require knowledge about an application.

The functions listed in section Function listing are available for use at any time.

Function listing

const char * cti_version(void)

cti_version returns a string containing the current frontend library version in the form major.minor.revision.

const char * cti_error_str(void)

When a CTI frontend function returns in error, the cti_error_str function can be used to obtain a verbose error string. It returns a string containing the human parsable error message, or else "Unknown CTI error".

int cti_error_str_r(char *buf, size_t buf_len)

• buf: Provided buffer to write the error string to.

• buf_len: Length of the user provided buffer.

cti_error_str_r is a re-entrant version of cti_error_str. It allows a user specified buffer to be passed in versus using a static global buffer. If the error string is longer than the provided buffer, the string is truncated and null-terminated.

cti_wlm_type_t cti_current_wlm(void)

cti_current_wlm is used to obtain the detected WLM. CTI has built in heuristics to detect which WLM is in use on the system. Users can explicitly override automatic WLM detection at runtime by setting the CTI_WLM_IMPL environment variable defined by the macro CTI_WLM_IMPL_ENV_VAR. See common_tools_fe.h for more information.

const char * cti_wlm_type_toString(cti_wlm_type_t wlm_type)

wlm_type: The cti_wlm_type_t to describe.cti_wlm_type_toString is used to obtain a human readable string representation of a cti_wlm_type_t.

char * cti_getHostname(void)

cti_getHostname is used to determine an externally-accessible hostname or IP address for the current node. This is the hostname of the network interface that can open socket connections between the login node and compute node. This is useful on systems where multiple network interfaces make a standard gethostname(2) call from glibc ambiguous.

int cti_setAttribute(cti_attr_type_t attrib, const char *value)

• attrib: attribute to modify.

• value: attribute specific value to set.cti_setAttribute is used to modify internal CTI configuration values. See common_tools_fe.h for a full accounting of attrib=value options that are available.

  const char * cti_getAttribute(cti_attr_type_t attrib)

• attrib: The requested cti_attr_type_t to obtain the current value.

cti_getAttribute is used to obtain the current value of the requested attribute. See common_tools_fe.h for a full accounting of available attribute options.

Applications

To use the majority of CTI functionality, a tool developer must first launch a new application under CTI control or register an already running application. When launching, CTI can also hold an application at a startup barrier before main. This allows the developer to launch tool daemons or stage files that are expected to be present before the application begins execution.

Upon successful launch or attach, a cti_app_id_t handle is returned. This opaque identifier is used for all application-specific functionality in CTI. The validity of an application handle can be determined using cti_appIsValid. An application handle is considered valid until the application exists (either normally/abnormally), or cti_deregisterApp is called. Signals can be sent to an application via the cti_killApp function.

Application launch

The cti_launchApp function is used to programmatically launch an interactive application. This replaces the manual fork/exec of launch commands such as aprun, srun, or mpiexec. CTI assumes a node allocation has been previously acquired, or nodes are marked as interactive, making compute resources available to the caller before use.

The application launcher employed is automatically detected by CTI. This logic is based on CTI detection of the workload manager (WLM) in use. See cti_current_wlm in the General frontend functions section for more info. A custom launcher can be explicitly specified with the CTI_LAUNCHER_NAME environment variable defined by the macro CTI_LAUNCHER_NAME_ENV_VAR.

cti_app_id_t cti_launchApp(const char * const   launcher_argv[],
                           int                  stdout_fd,
                           int                  stderr_fd,
                           const char *         inputFile,
                           const char *         chdirPath,
                           const char * const   env_list[])

• launcher_argv: A null-terminated list of arguments to pass directly to the launcher. It is the caller's responsibility to ensure that valid launcher_argv arguments are provided for the application launcher. The caller can use the cti_current_wlm function to determine which launcher is used by the system. Note that launcher_argv[0] must be the start of the actual arguments passed to the launcher, and not the name of launcher itself.

• stdout_fd: File descriptor in which to redirect stdout, or -1 if no redirection should take place.

• stderr_fd: File descriptor in which to redirect stderr, or -1 if no redirection should take place.

• inputFile: The pathname of a file in which to redirect stdin, or NULL to redirect /dev/null to stdin.

• chdirPath: The path in which to change the current working directory before launching the application, or NULL to use the existing current working directory.

• env_list: A null-terminated list of strings of the form "NAME=value" to set NAME in the application environment to value.

Upon success a non-zero cti_app_id_t is returned. On error, 0 is returned.

Application launch with barrier

A tool may require attaching onto an application before it begins execution, as is the case of a debugger, or bootstrapping itself early on. CTI supports an application launch variant cti_launchAppBarrier where the target application is held at a startup barrier before main. The cti_launchAppBarrier function takes the same arguments and has the same return value as cti_launchApp described in the Application launch section.

When a tool is ready to release the application from the startup barrier, it calls cti_releaseAppBarrier. This allows the application to continue normal execution.

int cti_releaseAppBarrier(cti_app_id_t app_id)

• app_id: The cti_app_id_t of the application launched via cti_launchAppBarrier.

Application attach

It is possible to use the CTI daemon and file transfer facilities with applications that were not started under direct control of CTI. In that case, there is no barrier equivalent as the application is already executing.

Registration of an existing app is largely specific to the WLM implementation. For example, an MPIR based launcher might require a pid_t of the application launcher process to which it is attached via ptrace. Alternative mechanisms besides MPIR are also available to exercise similar capabilities. For that reason, CTI uses a WLM specific identifier when possible. For example, registering an application with a Slurm based WLM requires two identifiers, jobid and stepid.

Because there is no one universal way to register existing applications with CTI, the different mechanisms are implemented as WLM-specific extensions. These are documented in the section WLM-specific Extensions.

WLM-specific Extensions

Most workload managers provide implementation-specific functionality. The most common example is in the attach case; each workload manager uses a different form of job identification to determine which application to attach. See section Application attach for more information.

CTI provides a generic extensible interface to add additional workload manager-specific functionality. To determine which workload manager is in use and thus which WLM extensions to call, use cti_current_wlm. See the General frontend functions section for more information.

See common_tools_fe.h for a list of all available WLM extensions.

Below is an example of attaching to a SLURM job using the CTI WLM extensions interface:

// Defined in common_tools_fe.h:
typedef struct {
    cti_app_id_t (*registerJobStep)(uint32_t job_id, uint32_t step_id);

    // Other SLURM operations...
} cti_slurm_ops_t;

// Application code
assert(cti_current_wlm() == CTI_WLM_SLURM);
cti_slurm_ops_t *slurm_ops = NULL;
assert(cti_open_ops(&slurm_ops) == CTI_WLM_SLURM);
cti_app_id_t const app_id = slurm_ops->registerJobStep(job_id, step_id);

Application facilities {#application-facilities .AlphaHeading2}

Once an application is registered, whether by launch or attach mechanisms, a variety of useful runtime facilities are available. These include querying application layout information, launching remote tool daemons on compute nodes, along with transferring files, binaries, libraries, and applicable dynamic library dependencies to a file system location accessible on the compute node.

Most runtime functions require an associated instance of cti_app_id_t to be provided, which is the application ID returned by the launch/attach described in the Application Lifetime section.

Function listing

int cti_getNumAppPEs(cti_app_id_t app_id)

Returns the number of processing elements (PE) in the application associated with the app_id. A PE represents an MPI rank for MPI based programming models.

int cti_getNumAppNodes(cti_app_id_t app_id)

Returns the number of compute nodes allocated for the application associated with the app_id.

char** cti_getAppHostsList(cti_app_id_t app_id)

Returns a null-terminated array of strings containing the hostnames of the compute nodes allocated by the application launcher for the application associated with the app_id.

cti_hostsList_t* cti_getAppHostsPlacement(cti_app_id_t app_id)

Returns a cti_hostsList_t containing entries that contain the hostname of the compute nodes allocated by the application launcher and the number of PEs assigned to that host for the application associated with the app_id.

File transfer support

A common requirement for tools is the ability to launch tool daemons alongside application ranks on compute nodes. This includes access to dependencies such as shared libraries or configuration files. CTI aims to provide an extensible interface that operates under many different constraints. A tool typically isn't concerned where a dependency resides on the file system. Rather, it cares that the dependency is accessible in a performant way.

For example, CTI aims to provide an interface that can cope with HPC systems that either have, or lack, a performant parallel file system. This may require co-locating the dependencies onto the compute nodes directly. It should also have the ability to provide system specific optimizations that prevent redundant transfer of dependencies already available via a parallel file system. All of this is achieved in a way that is transparent to the caller.

CTI manages unique storage locations via the paired concepts of sessions and manifests. These are described in following sections.

Sessions

The concept of a session allows CTI to manage different file system locations to which a tool daemon is guaranteed to have read/write access. A session represents a unique storage location where dependencies can be co-located, new files can be written, and is guaranteed to be cleaned up after the session/application exit.

Depending on tool need, multiple tool daemons can share the same session, or be isolated into different sessions. A session is always associated with an application via a cti_app_id_t. This is because a session must be associated with a file system location that may be unique to each compute node. This requires an associated application to describe this set of compute nodes.

The unique storage location of a session may be a parallel file system, or it may be a temporary storage location such as /tmp. The choice of where the storage location resides is implementation specific. CTI automatically creates unique directories in the base file system to create logical isolation between different sessions. This way, multiple tools can co-locate dependencies and run tool daemons concurrently without worry of clobbering file system locations.

Creation of the storage location associated with a session is deferred until a manifest (described in the Manifests section) is shipped or the tool daemon associated with the manifest is launched. A session has child directories for different dependencies: /bin for binaries, /lib for libraries, and /tmp for temporary storage. The TMPDIR environment variable of a tool daemon process will contain the associated session's /tmp location. Likewise LD_LIBRARY_PATH and PATH will point to the /bin and /lib location of the session respectiviely.

Function listing

cti_session_id_t cti_createSession(cti_app_id_t app_id)

• app_id: Application handle for a session.

A session is created with cti_createSession. This returns a cti_session_id_t session identifier for use with other interface calls. The validity of a session identifier can be determined using cti_sessionIsValid. A session is automatically invalidated if the associated cti_app_id_t becomes invalid.

int cti_destroySession(cti_session_id_t sid)

• sid: Session handle to destroy.

A session is destroyed via cti_destroySession. This will terminate every tool daemon associated with the session handle and remove the unique storage location if it was created. Tool daemons are terminated by sending a SIGTERM to the daemon process followed by a SIGKILL after 10 seconds. Upon completion, the session identifier becomes invalid for future use.

Manifests

Once a unique storage location is specified through the creation of a session, dependencies can be made available to it. This is achieved by generating a manifest and populating it with a list of files. A manifest is always associated with an owning session identifier. Sessions keep track of dependences previously made available to compute nodes. When a manifest is made available to a session, only those dependencies which are not already accessible to the session are co-located. This avoids redundant shipping of dependencies.

Function listing

cti_manifest_id_t cti_createManifest(cti_session_id_t sid)

• sid: Session id for the manifest.

A new manifest is created with cti_createManifest. This returns a cti_manifest_id_t manifest identifier for use with other interface calls. A manifest is automatically invalidated if the owning session becomes invalid. Dependencies contained within a manifest are not available to the session until a call is made to cti_sendManifest or cti_execToolDaemon. Once a manifest has been made available to the session, it is finalized and invalid for future modification. The validity of a manifest identifier can be determined using cti_manifestIsValid.

int cti_addManifestBinary(cti_manifest_id_t mid, const char *fstr)

• mid: The manifest id to which to add the dependency.

• fstr: The name of the binary to add to the manifest. This can either be the full path name of the binary or the base name of the binary in which case PATH is searched.

cti_addManifestBinary is used to add a program binary to a manifest. If the program binary is dynamically linked, its shared library dependencies will be automatically detected and added to the manifest. If the binary uses dlopen to open shared library dependencies, those libraries need to be added explicitly by calling cti_addManifestLibrary. This call is primarily for cases where a tool daemon launched via cti_execToolDaemon needs to fork/exec another program binary. This binary will be found in PATH and any shared library dependencies will be found in LD_LIBRARY_PATH of the environment of a tool daemon process.

If a shared library dependency is not available on the compute node and needs to be collocated, CTI is able to handle naming collisions across library names. CTI does this automatically via use of unique directories created under the session's /lib along with setting an appropriate LD_LIBRARY_PATH for the tool daemon(s). The same is not true for binaries or files; only unique binaries and files can be added to a session.

int cti_addManifestLibrary(cti_manifest_id_t mid, const char *fstr)

• mid: The manifest id to which to add the dependency.

• fstr: The name of the shared library to add to the manifest. This can either be a full path name, or the base name of the library. If a base name is specified, a search of library lookup paths will be conducted.

  int cti_addManifestLibDir(cti_manifest_id_t mid, const char *fstr)

• mid: The manifest id to which to add the dependency.

• fstr: The full path name of the directory to add to the manifest and make available within the /lib directory of the session.

cti_addManifestLibDir is used to add every library contained within a directory to the manifest. This is useful for programs that dlopen many dependencies. The directory structure will be preserved and found within the /lib directory of the session.

int cti_addManifestFile(cti_manifest_id_t mid, const char *fstr)

• mid: The manifest id to which to add the dependency.

• fstr: The full path name of the file to add to the manifest.

cti_addManifestFileis used to add an ordinary file to a manifest.

Object lifetime

There is an explicit ownership hierarchy defined within CTI. The topmost object is an application, represented by cti_app_id_t. The next object is a session, represented by a cti_session_id_t. At the bottom is a manifest, represented by a cti_manifest_id_t. Applications own sessions, which in turn own manifests. An important characteristic of CTI to recognize is this ownership definition. An application can own one or more session(s), and a session can own one or more manifest(s). That way, if the lifetime of an application ends, all owned sessions are invalidated, and internal data structures cleaned up. Likewise, if the lifetime of a session ends, all owned pending manifests are invalidated, and internal data structures cleaned up.

When invalidating a session via cti_destroySession, any tool daemons started within that session will also be killed. This behavior can be bypassed by calling cti_deregisterApp without explicitly calling cti_destroySession. This is useful for tools which are interested in bootstrapping their tool daemons from a login node without keeping a frontend presence alive. If a tool frontend exits without calling cti_deregisterApp or cti_destroySession, all launched applications and tool daemons will be killed.

There is no explicit way in the interface to invalidate a manifest. Manifests are lightweight lists of files and don't require any management considerations. A pending manifest has no impact on other manifests, state consideration happens only after a manifest has been made available to a session.

Tool daemons

Once a session is established and a manifest is created, tool daemon(s) can be launched onto the compute nodes associated with the session's application. CTI will launch a single tool daemon process onto every compute node associated with the application. It is up to the tool developer to fork/exec additional tool daemons if necessary, or exit if tool daemons need to execute only on a subset of compute nodes.

A manifest is required to be provided as part of a tool daemon's launch even if no other dependencies are required (i.e. the manifest is empty). Association with an application is made with the manifest argument: the manifest has an owning session which contains a list of already staged dependencies, and the session has an owning application to determine which nodes tool daemons need to be started.

CTI will conduct setup of the tool daemon environment before calling exec. This includes steps like the following:

• The tool daemon will have any binaries that have been made available to the session found within PATH.

• Shared libraries will be found within LD_LIBRARY_PATH.

• TMPDIR will point at the session specific /tmp location and is guaranteed to have read/write access.

• A null-terminated list of environment variables can be provided to set tool daemon specific environment. The specified argument array is provided to each tool daemon process.

• Any implementation specific tasks will be conducted for a particular system

By default, tool daemon processes will have their stdout/stderr redirected to /dev/null. This can be overridden by use of the CTI_DEBUG and CTI_LOG_DIR (see common_tools_shared.h for more information). This allows a tool daemon to write debug logs to a known location in a parallel file system. There will be one file per compute node with names correlating to each compute node number. The cti_setAttribute interface an also be used to define the logging behavior.

int cti_execToolDaemon( cti_manifest_id_t   mid,
                        const char *        fstr,
                        const char * const  args[],
                        const char * const  env[])

• mid: The manifest id to which to add the tool daemon binary.

• fstr: The name of the tool daemon binary. This can either be the full path name of the binary or the base name of the binary in which case PATH is searched.

• args: Null-terminated list of arguments to pass to the tool daemon. args[0] should be the first argument, not the name of the tool daemon binary.

• env: Null-terminated list of environment variables to set in the environment of the tool daemon process. Each variable setting should have the format envVar=val.

Backend library

Once a tool daemon is launched, the CTI backend library is available for use. This interface is defined in the common_tools_be.h header file. The libcommontools_be library should be linked into the tool daemon binary launched with cti_execToolDaemon. The backend library is used for determining node local information about the associated application. This information can be things like the logical PE ranks located on the node, the PID(s) of all application processes on the node, or filesystem layout of the session directory.

A subset of available functions is listed below.

Function listing

cti_wlm_type_t cti_be_current_wlm(void)

Returns the WLM in use by the application.

cti_pidList_t * cti_be_findAppPids(void)

Returns a cti_pidList_t containing the mapping of PE pid_t to logical PE rank.

char * cti_be_getNodeHostname()

Returns the hostname of the node.

int cti_be_getNodeFirstPE(void)

Returns the first logical PE number that resides on the node.

int cti_be_getNodePEs(void)

Returns the number of PE's on the node.