Last Updated: Nov 05, 2019
The HPSS connector can be used for accessing and sharing of data on an HPSS storage system. The connector is available as an add-on subscription to organizations with a Globus Standard subscription - please contact us for pricing.
This document describes the steps needed to install an endpoint and the HPSS connector needed to access the storage system. This installation should be done by a system administrator, and once completed, users can use the endpoint to access HPSS storage via Globus to transfer, share and publish data on the system.
In order for the HPSS DSI to function properly on the HPSS client node, please verify the following items.
Recently discovered issues and workarounds will be documented in the GitHub repository prior to inclusion in this document. See the repo Readme for details.
When a user accesses HPSS via GridFTP, home directory lookups and translations between usernames and user IDs are performed using the OS Name Service (ex /etc/passwd, ldap, nis, etc). The HPSS password file (ie HPSS_UNIX_AUTH_PASSWD) is not used by Globus. This has a direct impact on authentication and file access. Verify that HPSS users have the same UID on the local system and within HPSS. For example, given an HPSS user
hpssuser1, the UID and GID returned from the following command should match the UID and GID of the same account within HPSS:
$ getent passwd hpssuser1 hpssuser1:x:12345:1000:HPSS User:/home/hpssuser1:/bin/bash
Globus uses the account
hpssftp to access HPSS initially then changes user ID to the authenticated HPSS user (ie
hpssuser1). This removes the need to maintain per-user keytab files on the HPSS client node. However, this requires that the Globus process have access to the
hpssftp keytab entry during the authentication phase which runs under the authenticating user’s UID.
Assuming the keytab for
hpssftp is stored in
$ chmod 644 /var/hpss/etc/hpss.keytab
HPSS installations configured for Kerberos authentication must also allow non privileged users write access to HPSS temporary kerberos ticket cache, typically
$ chmod 1777 /var/hpss/cred
hpssftpkeytab file must not be exposed to unprivileged users. Prevent local shell access by non privileged HPSS users (ex. PAM).
As a non privileged HPSS user on the local node, verify that the local account is able to authenticate successfully to HPSS. For example:
$ /opt/hpss/bin/scrub -a krb5 -p hpssftp -k -t /var/hpss/etc/hpss.keytab scrub> quit
As a non privileged HPSS user, log into HPSS and perform some basic directory and file operations. Unlike the previous step, make sure these operations are performed as a non privileged user:
$ /opt/hpss/bin/scrub /hpss/home/testuser1 scrub> mkdir testdir scrub> rmdir testdir scrub> open testfile wc File created using COS 1 (Small File COS) scrub> write 5k .done (144.981 KB/sec) scrub> close scrub> unlink testfile scrub> quit
The HPSS DSI is compatible with the following Linux distributions:
The Globus Connect Server Installation Guide provides detailed documentation on the steps for installing and configuring a Globus endpoint. The HPSS DSI should be used with the latest version of Globus Connect Server 4.x.
This connector has been verified against HPSS versions 7.3, 7.4 and 7.5. Building HPSS is beyond the scope of this guide though you should have a working HPSS installation. The DSI requires either a full HPSS build or a
clnt HPSS build.
The following HPSS files located in
/var/hpss/etc are known to be required for operation of the HPSS DSI:
hpss.keytab (or hpss.unix.keytab)
These HPSS issues severely impact performance so the patches are highly recommended.
BZ2819 - PIO 60 second delay impacts small file performance. There is a small percentage chance that, after a transfer completes, HPSS PIO will wait 60 seconds before informing the client that the transfer has completed. This fix has been implemented in 7.3.3p9, 7.3.4, 7.4.1p1 and 7.4.2.
BZ2856 - Enabling HPSS_API_REUSE_CONNECTIONS returns address already in use. This one sets a limit on how many active connections we can have. GridFTP and HPSS make considerable use of ephemeral TCP ports. Quick, successive file transfers can lead the system to run out of available ports. There is no fix for this HPSS issue at this time. The number of ephemeral ports can be increased and the amount of time a socket spends in timed wait can be decreased to help avoid this issue.
BZ7772 - PIO 5 second delay impacts small file performance. There is a high percentage chance that, after a transfer completes, HPSS PIO will wait 5 seconds before informing the client that the transfer has completed. This greatly impacts the performance of file retrieves and checksum operations. This fix has been implemented in 7.5.3+.
BZ7883 - Prevents successful transfers of files over 4GiB on HPSS versions 7.5.2+. Due to what appears to be a transfer length calculation error, transfer of files larger than 4GiB generate an EIO error at the 4GiB mark and the transfer terminates. This bug impacts all HPSS clients using the HPSS PIO interface. Upgrade to HPSS 7.5.2u5 / HPSS 7.5.3u1 to resolve this issue.
GridFTP installations benefit from and take full advantage of classes of service that use fixed length classic style allocation. In short, you’ll get the best performance from the GridFTP interface (actually any HPSS interface) if the segment count is below 32.
HPSS has multiple disk/tape allocation algorithms used to allocate space for incoming data. Fixed length allocation gives you equal size chunks to store data in. This was deemed wasteful because the last block was most certainly never filled. Variable length allocation was created to solve this problem; it will give you increasingly larger segments as data is stored and truncates the last block. This is a win for most situations when HPSS is unsure how much data is to be stored for the given file.
Using either of these allocation mechanisms (any variable length allocation or fixed w/o knowing the file size), HPSS is free to continue to allocate segments until all the data is stored. This has a definite performance impact because internally HPSS retrieves data in 32-segment chunks. This means when you request a file from HPSS, internally it breaks it up into multiple transfers, each of which is ⇐ 32 segments. Functionally, this is transparent to the client. In terms of performance, the client will see a high load followed by a pause followed by a high load, etc.
In order to avoid the performance hit, you can use fixed length allocation with segment counts < 32 and take advantage of the fact that any WELL-BEHAVED GridFTP client will inform HPSS of the size of the incoming file before the transfer begins. In fact, the DSI is designed to require this. If a GridFTP client is not well behavad, the DSI will act as though a zero length transfer is about to occur and will handle it as such. So you’ll know if the client is not doing the right thing.
As of version 2.9, the HPSS Connector is installed from RPM instead of building from source. Because of this, several changes to the system for installation of previous versions need to be reversed so that they do not conflict with the RPM installation. This includes:
Make sure this file does not exist. Also remove any other files within the same directory that supply configuration options for
Remove this file if it exists so that the GridFTP server can find the new HPSS DSI in /usr/lib64/. Be sure to run
ldconfigto update system paths.
Delete or move this file to save it to avoid conflict with the RPM-managed version of the file.
The HPSS Connector is installed by RPM in order to simplify installation and enforce version requirements with requisite software. Unfortunately, Globus does not have access to all HPSS and RHEL version combinations to provide RPMs for all installations. If an RPM is not available for your version, instructions below explain how to obtain the source and generate the RPM.
Visit the release page and find the latest release. Note that release candidates are also available from this page and are indicated with the
Pre-release badge and have a
Release Candidate name. For production use, make sure to choose the most recent official release which will have the
Latest Release badge.
Go to the
Assets section under the latest release and you’ll see that you have two choices for installation. When possible, RPMs for supported platforms will be available for your convenience. If an RPM is not available, you’ll need to create it using the instructions below.
The RPM has a naming scheme of:
This package is for HPSS version 7.5.X
This is the first release of this connector version 2.9
This package is for RHEL 6.X.
Download the RPM and matching
.asc file which will allow you to verify that the RPM has not changed since creation. Using a recent version of gpg (>= 2.0), import the public key used for signing:
$ gpg --keyserver hkp://keys.gnupg.net --recv-keys 1EA106A24003C353 gpg: requesting key 4003C353 from hkp server keys.gnupg.net gpg: key 4003C353: "Jason Alt <firstname.lastname@example.org>" imported gpg: Total number processed: 1 gpg: imported: 1
And then verify the downloaded RPM:
$ gpg --verify globus-gridftp-server-hpss-7.5-2.9-1.el7.x86_64.rpm.asc globus-gridftp-server-hpss-7.5-2.9-1.el7.x86_64.rpm gpg: Signature made Wed 06 Nov 2019 09:45:26 PM UTC using RSA key ID 4003C353 gpg: Good signature from "Jason Alt <email@example.com>" gpg: aka "Jason Alt <firstname.lastname@example.org>" gpg: WARNING: This key is not certified with a trusted signature! gpg: There is no indication that the signature belongs to the owner. Primary key fingerprint: C36C 826C 18ED 73C3 38DC FA53 1EA1 06A2 4003 C353
And finally, install the downloaded RPM using YUM:
$ sudo yum install ./globus-gridftp-server-hpss-7.5-2.9-1.el7.x86_64.rpm
If an RPM is not available for your platform, you will need to build it for your self. Under
Assets, download and un-tar the assest name
Install the development prerequisite libraries:
$ sudo yum install openssl-devel \ globus-gridftp-server-devel \ hpss-lib-devel
From within the un-tar’ed source code directory, use the
release make target to build the RPMs for your platform:
$ make -f Makefile.bootstrap release
Once the RPM is built, install it:
$ sudo yum install ./globus-gridftp-server-hpss-7.5-2.9-1.el7.x86_64.rpm
/var/hpss/etc/gridftp_hpss_dsi.conf in the source directory for any changes you may wish to make for your site. You will likely leave most of these options commented out to use their default values.
- LoginName <user>
(optional) This is the HPSS service user used to initially authenticate with HPSS. GridFTP requires a privileged user with control permission on the core server’s client interface in order to log into HPSS and then change its credentials to that of the connecting user. Defaults to
hpssftpwhich is also handled special by HPSS with regards to gate keeper operations.
- AuthenticationMech [unix|krb5|gsi|spkm]
(optional) Defines the type of authentication that the DSI will perform when logging into HPSS. Note that this is not the authentication mechanism the GridFTP users will use; they always use GSI. Defaults to HPSS_API_AUTHN_MECH or HPSS_PRIMARY_AUTHN_MECH.
- Authenticator [auth_keytab|auth_keyfile|auth_key|auth_passwd][:<file>]
(optional) Defines the location of credentials to be used by the DSI to authenticate to HPSS as
LoginName. Defaults to HPSS_PRIMARY_AUTHENTICATOR. When this option points to a file, that file’s contents must be accessible to
For unix authentication, you can put the
LoginNameaccount credentials into its own file using
Authenticatorto that file instead of giving the GridFTP process read access to the target of HPSS_PRIMARY_AUTHENTICATOR.
For sites using kerberos authentication with HPSS, you’ll need to create a kerberos keytab file using the kerberos utility
ktutilif you wish to seperate the
LoginNamecredentials from the target of HPSS_PRIMARY_AUTHENTICATOR.
- QuotaSupport [on|off]
(optional) This option is deprecated and will be removed in a future release.
- UDAChecksumSupport [on|off]
(optional) Causes checksums to be stored within UDAs so that the checksum can be recalled later without bringing the file back from tape. It is recommended that you set this option to
onto avoid uncessary tape recalls. The default is
Recent changes to make use of the async stage request API for HPSS in order to
avoid inundating the core server with duplicate stage requests has exposed a
deficiency for the DSI use case of HPSS. The HPSS async stage API expects the
call to be available long term in order to receive stage completion messages.
However, the GridFTP/DSI use case is a short-lived transient environment; the
GridFTP process can not wait minutes/hours/days for stage completion messages.
Users of DSI versions 2.6+ will see the impact as a
in the HPSS GUI console. The warning is innocuous and can be ignored. IBM is
aware of this issue and a change request has been created.
As a work around, users of 2.6 should update to 2.7 and all users of 2.7+ can
blackhole sync method. This configures
nc (netcat) to listen for
stage completion messages intended for the DSI and discard whatever it receives.
nc should be launched on a highly-available server reachable by the HPSS core
servers (preferably run it directly on the core servers). Choose a port to use
for receiving callback notifications on and run this command:
admin@hpss-core $ nc -v -v -k -l <port>
nc is running, add this to
/etc/gridftp.d/hpss_issue_35 on the GridFTP
nodes running the HPSS DSI:
These options are generally set in /var/hpss/etc/env.conf and affect operation of data transfers. Some of these options may be required depending upon your configuration.
This option selects the network interface used for data transfers between the Globus services you are configuring and the HPSS mover machine(s). If this is unset, data transfers use the default network interface. This option is generally necessary on multihomed nodes. It should be set on the node running GridFTP.
This controls the amount of time before a mover process will stop waiting for data from the Globus service in order to reclaim network resources. Default is 15 minutes. In very large file transfers, it is possible that movers may timeout before the transfer reaches data offsets which those movers are responsible for. This option is set on the mover nodes.
Kerberos must be configured for access to the proper Kerberos realm that contains HPSS. This file is usually kept in /etc/krb5.conf. You may need to enable the allow_weak_crypto option in the [libdefaults] section if the DSI module can not talk to the HPSS servers.
After completing the installation, you should do some basic transfer tests with your endpoint to ensure that it is working. We document a process for basic endpoint functionality testing here.
Below are some common issues encountered while using the Globus Transfer service with an endpoint running the HPSS connector along with possible resolutions to each problem.
Login Failure: No such file or directory
This error message indicates that hpss_LoadDefaultThreadState() has returned
ENOENT causing the login procedure to fail. This is occurs when the UID of the authenticating user as known to the GridFTP process does not match the user’s ID as known by HPSS. See Local user accounts must match user accounts in HPSS.
Command Failed: Error (login) Endpoint: xxxx Server: xxxx Message: Login Failed --- Details: 530-Login incorrect. : GlobusError: v=1 c=PATH_NOT_FOUND\r\n530-GridFTP-Errno: 2\r\n530-GridFTP-Reason: System error in hpss_LoadDefaultThreadState()\r\n530-GridFTP-Error-String: No such file or directory\r\n530 End.\r\n
Login Failure: Operation not permitted
This error message indicates that hpss_SetLoginCred() failed with
EPERM during the login procedure. This step in the login process accesses the keytab defined in
AuthenticationMech so that the DSI can connect to HPSS as user
LoginName. The error value indicates that the GridFTP process was unable to access the keytab file. See hpssftp credentials must be accessible by local unprivileged accounts.
Command Failed: Error (login) Endpoint: xxxx Server: xxxx Message: Login Failed --- Details: 530-Login incorrect. : GlobusError: v=1 c=INTERNAL_ERROR\r\n530-GridFTP-Errno: 1\r\n530-GridFTP-Reason: System error in hpss_SetLoginCred()\r\n530-GridFTP-Error-String: Operation not permitted\r\n530 End.\r\n
Login Failure: Cannot access config file
The following error implies that /var/hpss/etc/gridftp_hpss_dsi.conf does not exist.
Error (login) Endpoint: XXX Server: XXX Message: Login Failed --- Details: 530-Login incorrect. : GlobusError: v=1 c=PATH_NOT_FOUND\r\n530-GridFTP-Errno: 2\r\n530-GridFTP-Reason: System error in Can not access config file\r\n530-GridFTP-Error-String: No such file or directory\r\n530 End.\r\n
Login Failure: Invalid argument
If you receive this message, it is likely that /var/hpss/etc/site.conf is invalid.
Error (login) Endpoint: XXX Server: XXX Message: Login Failed --- Details: 530-Login incorrect. : GlobusError: v=1 c=INTERNAL_ERROR\r\n530-GridFTP-Errno: 22\r\n530-GridFTP-Reason: System error in hpss_LoadDefaultThreadState()\r\n530-GridFTP-Error-String: Invalid argument\r\n530 End.\r\n
Transfer Error: Data channel authentication failed
When transfers of zero-length files fail frequently as shown below, it can generally be traced back to an invalid threads setting in
/etc/gridftp.d/hpss. The HPSS DSI must be configured with a
threads value of 2 or greater. See Configure GridFTP to Use the HPSS Connector for the proper configuration.
Error (transfer) Endpoint: XXX Server: XXX File: /~/zero_length_file Command: RETR ~/zero_length_file Message: Data channel authentication failed --- Details: 500-Command failed. : globus_xio: The GSI XIO driver failed to establish a secure connection. The failure occured during a handshake read.\r\n500-globus_xio: Operation was canceled\r\n500-globus_xio: Operation timed out\r\n500 End.\r\n
Transfer Error: Operation timed out
Large file transfers to/from HPSS tend to span multiple sets of HPSS mover processes. Each set is responsible for a large contiguous chunk of the file transfer. First set transfers offsets 0-N, second set transfers (N+1)-M, and so on. These mover sets are all initialized at the beginning of the transfer.
Any mover will timeout after MVR_CLIENT_TIMEOUT seconds (defaults to 15 minutes). If a mover set does not start the transfer within this timeout, the entire transfer aborts. This is an HPSS issue, not a DSI issue.
This error condition is usually obvious from the following errors issued in MVR_CLIENT_TIMEOUT seconds + 5 minute intervals. See MVR_CLIENT_TIMEOUT for more details.
2019-06-12 14:29:39 Error (transfer) Endpoint: XXXX HPSS Archive (e38ee901-6d04-11e5-ba46-22000b92c6ec) Server: XXXX:2811 Command: STOR ~/scratch_backups/XXXX Message: The operation timed out --- Details: Timeout waiting for response
2019-06-12 14:49:47 Error (transfer) Endpoint: XXXX HPSS Archive (e38ee901-6d04-11e5-ba46-22000b92c6ec) Server: XXXX:2811 File: /~/scratch_backups/XXXX Command: STOR ~/scratch_backups/XXX Message: Fatal FTP response --- Details: 451-GlobusError: v=1 c=INTERNAL_ERROR\r\n451-GridFTP-Errno: 5011\r\n451-GridFTP-Reason: System error in hpss_PIOExecute\r\n451-GridFTP-Error-String: \r\n451 End.\r\n
If errors occur, the following information can be collected in order to better diagnose the issue.
/etc/gridftp.d/hpss_debug and include the following values which will cause the HPSS API to send verbose output to the file specified by
$HPSS_API_DEBUG 7 $HPSS_API_DEBUG_PATH <path>
Make sure the file exists prior to collecting debug information and is writable by the end user:
$ sudo touch <path> $ sudo chmod 600 <path>
The server begin to write out debug information on the next connection to the GridFTP server.