[z/OS]

Use of storage class memory with shared queues

The use of storage class memory (SCM) can be advantageous when used with IBM® MQ for z/OS® shared queues.

Important: IBM z16 is planned to be the last generation of IBM Z® to support the use of Virtual Flash Memory (also known as Storage Class Memory, or SCM) for Coupling Facility images. For more information see: IBM Z and IBM LinuxONE 4Q 2023 Statements of Direction.

As an alternative, you should either use larger structures or offload messages to SMDS.

The z13, zEC12, and zBC12 machines allow the installation of Flash Express cards. These cards contain flash solid-state drives (SSD). After installation, flash storage from the cards can be allocated to one or more LPARs where it is typically known as SCM.

SCM sits between real storage and direct access storage device (DASD) in terms of both I/O latency and cost. Because SCM has no moving parts, it exhibits much lower I/O latencies than DASD.

SCM is also much cheaper than real storage. As a result, a large amount of storage can be installed for a relatively low cost; for example, a pair of Flash Express cards contains 1424 GB of usable storage.

These characteristics mean that SCM is useful when a large amount of data must be taken from real storage in a short period of time, because that data can be written to SCM much quicker than it can be written to DASD. This specific point can be very useful when using coupling facility (CF) list structures containing IBM MQ shared queues.

Why list structures fill up

When a CF structure is defined, it is configured with a SIZE attribute that describes the maximum size of the structure. Because CF structures are always permanently resident in real storage, the sum of the SIZE attributes of the structures that are defined on a CF should be less than the amount of the real storage allocated to the CF.

As a result, there is a constant pressure to keep the SIZE value of any given structure to the minimum value possible, so that more structures can fit into the CF. However, ensuring structures are large enough to achieve their purpose can result in a conflicting pressure, because making a structure too small means that it might fill up, disrupting the applications or subsystems making use of it.

There is a strong need to accurately size a structure based on its expected usage. However, this task is difficult to do because workloads can change over time and accounting for their fluctuations is not easy.

IBM MQ shared queues use CF list structures to store messages. IBM MQ calls CF structures, which contain messages and application structures.

Application structures are referenced using the information stored in IBM MQ CFSTRUCT objects. When a message that is smaller than 63 KB is put on a shared queue, the message is stored entirely in an application structure as a single list entry, and zero or more list elements.

Because IBM MQ shared queues use list structures, the pressures described also affect shared queues. In this case, the maximum number of messages that can be stored on a shared queue is a function of the:
  • Size of the messages on the queue
  • Maximum size of the structure
  • Number of entries and elements available in the structure
Because up to 512 shared queues can use the same structure, and effectively compete for entries and elements, this complicates matters even further

IBM MQ queues are used for the transfer of data between applications, so a common situation is an application putting messages to a queue, when the partner application, which should be getting those messages, is not running.

When this happens the number of messages on the queue increase over time until one or more of the following situations occur:
  • The putting application stops putting messages.
  • The getting application starts getting messages.
  • Existing messages on the queue start expiring, and are removed from the queue.
  • The queue reaches its maximum depth in which case an MQRC_Q_FULL reason code is returned to the putting application.
  • The structure containing the shared queue reaches its maximum size, or the CF containing the structure runs out of available storage. In either case, an MQRC_STORAGE_MEDIUM_FULL reason code is returned to the putting application.
In the last three cases the queue is full. At this point the putting application has a problem because there is nowhere for its messages to go. The putting application typically solves this problem by using one or more of the following solutions:
  • Repeatedly retry putting the message, optionally with a delay between retries.
  • Put the messages somewhere else, such as a database or a file. The messages can be accessed later and put to the queue as normal.
  • Discard the message if it is nonpersistent.

However, for some classes of applications, for example those with a large volume of incoming messages, or no access to a file system, these solutions are not practical. There is a real need to ensure that queues never, or are extremely unlikely to, fill up in the first place and this is especially pertinent to shared queues.

SMDS and offload rules

The offload rules introduced in IBM WebSphere® MQ 7.1 provide a way of reducing the likelihood of an application structure filling up.

Each application structure has three rules associated with it, specified using three pairs of keywords:
  • OFFLD1SZ and OFFLD1TH
  • OFFLD2SZ and OFFLD2TH
  • OFFLD3SZ and OFFLD3TH
Each rule specifies the conditions that must be met for message data to be offloaded to the storage mechanism that is associated with the application structure. Two types of storage mechanisms are currently available:
  • Db2®
  • Group of Virtual Storage Access Method (VSAM) linear data sets, which IBM MQ calls a shared message data set (SMDS).

The following example shows the MQSC command to create an application structure named LIST1, using the DEFINE CFSTRUCT command.

This structure has the default offload rules in place, and uses SMDS as the offload mechanism. This means that when the structure is 70% full ( OFFLD1TH), all messages that are 32 KB or larger ( OFFLD1SZ) are offloaded to SMDS.

Similarly, when the structure is 80% full ( OFFLD2TH) all messages that are 4 KB or larger ( OFFLD2SZ) are offloaded. When the structure is 90% full ( OFFLD3TH) all messages ( OFFLD3SZ) are offloaded. 

DEFINE CFSTRUCT(LIST1)
CFLEVEL(5)
OFFLOAD(SMDS)
OFFLD1SZ(32K) OFFLD1TH(70)
OFFLD2SZ(4K) OFFLD2TH(80)
OFFLD3SZ(0K) OFFLD3TH(90)

An offloaded message is stored in the offload medium, and a pointer to the message is stored in the structure. While the offload rules reduce the chance of the structure filling up, by putting less message data in the structure as it runs out of storage, some data is still written to the structure for each message. That is, the pointer to the offloaded message.

Additionally, the offload rules come with a performance cost. Writing a message to a structure is relatively quick and is largely dominated by the time spent to send the request for the write to the CF. The actual writing to the structure is fast, happening at real storage speeds.

Writing a message to SMDS is much slower because it includes writing to the structure for the message pointer, and writing the message data to SMDS. This second write operation is done at DASD speed and has the potential to add latency. If Db2 is used as the offload mechanism the performance cost is much greater.