IBM Cloud Object Storage System : Packed Slice Storage€¦ · Chapter 1. Overview IBM Cloud Object Storage System ™ pr ovides various interfaces to stor e and r etrieve objects

IBM Cloud Object Storage System™

Version 3.14.3

Packed Slice Storage

DSNADM_PSS-20160318D

IBM

NoteBefore using this information and the product it supports, read the following information:

v The general information in Notices

v The information in Safety and environmental notices

v The information in the IBM Environmental Notices and User Guide (provided on a DVD)

This edition applies to IBM Cloud Object Storage System™ and is valid until replaced by new editions.

© Copyright IBM Corporation 2016, 2019.US Government Users Restricted Rights – Use, duplication or disclosure restricted by GSA ADP Schedule Contractwith IBM Corp.

Contents

Chapter 1. Overview . . . . . . . . . 1

Chapter 2. New capabilities . . . . . . 3

Chapter 3. Benefits . . . . . . . . . . 5

Chapter 4. Terminology and components 7

Chapter 5. Workflows . . . . . . . . . 9Writing an object . . . . . . . . . . . . . 9

Writing a slice to an existing bin . . . . . . . 9Writing a slice to a new bin . . . . . . . . 9Committing a slice . . . . . . . . . . . 9Finalizing a slice. . . . . . . . . . . . 10

Reading a slice . . . . . . . . . . . . . 10

Deleting a slice . . . . . . . . . . . . . 10Listing a slice. . . . . . . . . . . . . . 10Compacting a bin file . . . . . . . . . . . 11Improved performance with small objects . . . . 11Deploying to new storage pools only . . . . . . 11

Chapter 6. Deploying . . . . . . . . 13

Chapter 7. Operations . . . . . . . . 15

Index . . . . . . . . . . . . . . . 17

Trademarks. . . . . . . . . . . . . 19

Homologation statement . . . . . . . 21

© Copyright IBM Corp. 2016, 2019 iii

iv IBM Cloud Object Storage System™: Packed Slice Storage

Chapter 1. Overview

IBM Cloud Object Storage System™ provides various interfaces to store and retrieve objects on an IBMCloud Object Storage Slicestor® system.

Every object is converted into a set of system-addressable entities called data sources. For every object, ametadata data source and zero or more object data sources exists. Metadata data sources contain systemobject metadata, while object data sources contain user-supplied object data. If an object is relativelysmall, all of the object data will be embedded in the metadata object and no data object sources will begenerated. In addition, to support name listing operations, the system maintains additional data sourcesthat store a name index.

Each data source undergoes multiple transformations and additional data is added to guarantee datareliability and integrity. The result of these transformations are called slices: pieces of data sufficient torecreate the original data with bit-level accuracy as long as a defined threshold of slices out of the entiredefined IDA width are available. Slices are sent to IBM Cloud Object Storage Slicestor® nodes, whichpersist each slice to a specific disk for long-term data storage. The process by which a IBM Cloud ObjectStorage Slicestor® appliance organizes, persists, and retrieves slices on a given disk is called a backendstorage methodology.

Traditionally, all data source slices have been stored with the File Slice Storage (FSS) backend storagemechanism. File Slice Storage stores slices as one file per slice. While this generally works well, there aredisadvantages for smaller slices (metadata data source slices):v Significant storage overhead per slicev Inefficient structure for listing slicesv Low throughput and IOPS

© Copyright IBM Corp. 2016, 2019 1

2 IBM Cloud Object Storage System™: Packed Slice Storage

Chapter 2. New capabilities

Release 3.4.0 introduced a new backend storage methodology called Packed Slice Storage (PSS). PackedSlice Storage works in conjunction with File Slice Storage; it stores metadata data source slices efficientlywhile other data source slices continue to be managed by File Slice Storage.v PSS is shared storage for all vaults deployed on a given Slicestor® Node.v PSS greatly improves performance in the storage and retrieval of small objects.



Chapter 3. Benefits

PSS offers the following benefits:v Use storage space more efficiently.v Read objects with greater efficiency: single seek in most cases.v Write objects with greater efficiency: data is appended to the end of one of a fixed number of files that

PSS maintains.v Improve IOPS and lowered latency. In most cases, no actual I/O occurs when trying to read

non-existent slices.v Reduce the burden on underlying file system.v Improve listing operations in both speed and system resources utilized.v Perform at the same level for long periods at a wide range of fills due to the reduced load on

underlying drives.



Chapter 4. Terminology and componentsPSS instance

A set of Bin Files on a single drive. One is deployed separately per drive. There are as manyinstances as there are active drives in a running Slicestor node.

Bin Files (128 per drive) The source for all data associated with a PSS instance. Bin files are used for all data-relatedoperations and to recover other components in disaster recovery situations. Each bin file containsa header, and zero or more bins.

Bin A container that stores zero or more slices that can be referenced individually in a bin file. Slicesare assigned to bins based on a slice name hash value.

Note: Potentially, multiple slices could share a bin, which is known as a hash collision. Thenumber of bins on disk is large enough to avoid frequent hash collisions.

In-memory pointersAll bin locations are stored in memory; these in-memory pointers are used when servicingnormal client IO. The amount of pointers that can be held depends on available system memoryat the time of storage pool creation.

Pointer file (one per disk) Every running PSS instance maintains a log-based journal file for bin locations written to PSS. Itis used to restore bin references to system memory during system startup only. It is not needed toperform operations. The pointer file is appended with bin references each time a slice is written.It is periodically compressed to remove old pointer values.

Catalog (one instance per disk) A database that holds extant slice names and their revisions. This database allows for efficientclient object slice listing operations.



Chapter 5. Workflows

Writing an objectWriting an Object to PSS works exactly like FSS. Writes are completed in three phases: write, commit, andfinalize.

Write Slice data is written to disk. The slice is now durable, but cannot be read back until it iscommitted.

Commit Data is made visible to client systems or APIs. An attempt to access this object will succeed. Thecommit phase starts once slices are written to a number of IBM Cloud Object Storage Slicestor®

nodes equal to the IDA Write Threshold.

Finalize If the object being written is an update of an existing object, the previous revision of the data isremoved from the system when it is finalized. The finalize phase starts once slices have beencommitted to a number of IBM Cloud Object Storage Slicestor® nodes equal to the IDA WriteThreshold.

Writing a slice to an existing bin

If the slice is being written to an existing bin:1. The slicename is hashed to determine bin file and bin number.2. Existence of an in-memory pointer indicates the bin in question already exists on disk.3. The bin location is retrieved from the in-memory pointers.4. The entire bin is read into system memory.5. The bin is modified to include the new slice data.6. The entire new bin is appended to the end of the bin file.7. The in-memory bin pointers are updated to reflect the new position of this bin.8. The Pointer File is updated with the new location of the bin.

Writing a slice to a new bin

About this task

You can write a slice to a new bin.

Procedure1. The slicename is hashed to determine bin file and bin number.2. Non-existence of an in-memory pointer indicates the bin in question does not already exist on disk.3. A new bin is created in memory containing the slice.4. The new bin is appended to the appropriate bin file.5. The in-memory bin pointers are updated to reflect the new position of this bin.6. The Pointer File is updated with the location of the new bin.

Committing a slice1. The slicename is hashed to determine bin file and bin number.2. The bin location is retrieved from in-memory pointers.


3. The bin header is updated in place on disk to reflect that the slice data has been committed.4. The Catalog is updated with the name of the slice data and its current revision.

Finalizing a slice

About this task

You can finalize a slice.

Procedure1. The slicename is hashed to determine bin file and bin number.2. The bin location is retrieved from in-memory pointers.3. The bin header is updated in place on disk to reflect that old revisions of the slice data have been

finalized.4. Old revision entries for the slice data are removed from the Catalog.

Reading a slice

About this task

A slice can be read.

Procedure1. The slicename is hashed to determine bin file and bin number.2. The bin location is retrieved from in-memory pointers.3. The bin is read from disk using said location.

Results

The slice is sent to the IBM Cloud Object Storage Accesser® node for reassembly into an object to be sentto the client system.

Deleting a slice

About this task

A slice can be deleted.

Procedure1. The slicename is hashed to determine bin file and bin number.2. The bin location is retrieved from in-memory pointers.3. The bin header is modified in place on disk to indicate the slice has been deleted.

Results

Deletes are committed and finalized like normal writes.

Listing a sliceAll slice names are stored in the Catalog. Any list operations request the slice names from the Catalogand return the requested values.


Compacting a bin fileSome write (if overwriting a bin is required) and delete operations remove references to the affected bin,making the contained data inaccessible. The unreferenced bin still occupies space in a bin file. Theprocess of removing unreferenced bins from bin files is called compaction. It is executed in the backgroundwithout disrupting regular storage functions. Compaction occurs more frequently when disk utilization isclose to 100% to reclaim unused space more quickly.

Improved performance with small objectsPSS has advantages over FSS as it uses a predictable access pattern to store and retrieve data.v A read operation requires a single seek to find the location of a bin containing the requested slice on a

disk.v A write operation always appends the content of a slice as part of an associated bin to a bin file based

on hash value for this slice name. If an affected bin already exists, it has to be first read, then modifiedwith the content of the new slice and then appended to the bin file. Checking for the existence of a binprior to a write operation does not require disk access or any I/O operations.

PSS exhibits far less performance degradation than FSS over the system lifetime. Read and writeoperation performance does not degrade due to number of objects. During compaction, performancecould degrade temporarily.

Deploying to new storage pools onlyPSS can be deployed on new storage pools only. All vaults sharing a PSS-enabled pool would haverelevant slices stored in PSS. There is no mechanism to grant PSS access on the vault level. PSS improvesperformance and improves disk space utilization when used in CSO mode with large number of objectsunder 10 MB in size. The impact is especially significant when system load is dominated by objectssmaller than 100 KB. PSS could improve latency even for large objects when storage pools have a varietyof object sizes.

Note: Do not enable PSS for pools exclusively used with SO vaults.

Note: For workloads that contain exclusively large object sizes, PSS does not offer any performancebenefits.

Chapter 5. Workflows 11


Chapter 6. Deploying

About this task

You can deploy the software.

Procedure1. Create a storage pool:

a. Click the Configure tab.b. Click the link Create Storage Pool in the Summary section.

Storage pools can have different capacity nodes. A model group, consisting of nodes that can beincluded in the same storage pool, must be selected. If nodes across multiple sites are available,the required site must be selected. If a storage pool will have nodes spanning multiple sites,balance the nodes across the sites. Not balancing the nodes can introduce read or write availabilityissues in the case of a single site outage.

2. Type a name for this storage pool in the Name field.3. Select an IDA width for this storage pool from the Width drop-down box.4. Select the Packed Storage radio button to enable packed slice storage for this storage pool.

Note: Contact support to determine if enabling packed slice storage will benefit current storageneeds.

Note: Packed storage can only be enabled when all IBM Cloud Object Storage Slicestor® nodes havebeen upgraded to 3.4.

5. Nodes may be selected either automatically (Suggest Devices) per the above rules, or manually. Toselect multiple devices:a. Select the first node.b. Hold down the B key.c. Select the last device.

6. Once the devices are identified, click Save to create the Storage Pool.CAUTION: If there is not an even distribution of Slicestor® Nodes across the available sites where theloss of one site would make Vaults either unusable or read-only, the Manager Web Interface displays aconfirmation dialog box that asks the operator if they accept the settings with the risks they present:The selected devices are not balanced evenly across sites. This could lead toread and write availability issues in the case of a site outage. Do you stillwish to continue?

You can click Cancel or OK to change or keep the settings.From the Configure Storage Pool page, click Change to rename this storage pool or Monitor > toshortcut to the monitor function for this storage pool. A different storage pool may be selected fromthe Storage Pool landing page.



Chapter 7. Operations

Quarantined disk events are typically generated with a reason code parameter.

The table below provides reason codes for packed storage.

Table 1. Reason codes

Code Reason

30 An unreadable pointer file. Contact support.

31 An unreadable catalog file. Contact support.

32 An invalid packed directory structure. Contact support.

33 Too many unreadable bin files - retention area is full. Contact support.

Example of new event message for packed slice storage:Disk in drive bay <bay> with S/N <serial> is quarantined, due to an unreadable pointerfile. Please contact Customer Support. Reason code is 30.



Index

Hhomologation statement 21

Ttrademarks 19



Trademarks

IBM®, the IBM logo, and ibm.com® are trademarks or registered trademarks of International BusinessMachines Corp., registered in many jurisdictions worldwide. Other product and service names might betrademarks of IBM or other companies. A current list of IBM trademarks is available on the web atCopyright and trademark information at www.ibm.com/legal/copytrade.shtml.

Accesser®, Cleversafe®, ClevOS™, Dispersed Storage®, dsNet®, IBM Cloud Object Storage Accesser®, IBMCloud Object Storage Dedicated™, IBM Cloud Object Storage Insight™, IBM Cloud Object StorageManager™, IBM Cloud Object Storage Slicestor®, IBM Cloud Object Storage Standard™, IBM Cloud ObjectStorage System™, IBM Cloud Object Storage Vault™, SecureSlice™, and Slicestor® are trademarks orregistered trademarks of Cleversafe, an IBM Company and/or International Business Machines Corp.

Other product and service names might be trademarks of IBM or other companies.


http://www.ibm.com/legal/copytrade.shtml


Homologation statement

This product may not be certified in your country for connection by any means whatsoever to interfacesof public telecommunications networks. Further certification may be required by law prior to making anysuch connection. Contact an IBM representative or reseller for any questions.



IBM®

Printed in USA

IBM Cloud Object Storage System : Packed Slice Storage€¦ · Chapter 1. Overview IBM Cloud Object Storage System ™ pr ovides various interfaces to stor e and r etrieve objects

Documents