VM65992: HIPERSOCKETS PERFORMANCE ISSUES ON SHORT BUSY

Obtain the fix for this APAR.

APAR status

• Closed as program error.

Error description

• Using a dedicated HiperSockets device in a virtual machine
  running Linux exploiting QDIO Enhanced Buffer State Management
  (QEBSM), may experience slow performance in a very highly
  contested LPAR to LPAR communication environment.  This is an
  environment where multiple HiperSockets data transmissions occur
  simultaneously to the same set of QDIO queues.  It's this
  contested environment which increases the likelihood of a
  HiperSockets device presenting a short busy condition on a
  Signal Adapter (SIGA) Instruction issued by a program (the Linux
  Guest in this case).  It is the occurrence of a HiperSockets
  short busy which opens an error window that may cause the
  problem to occur.
  Running with QIOASSIST OFF increases the likelihood
  of seeing this problem.
  

Local fix

Problem summary

• ****************************************************************
  * USERS AFFECTED: All customers using a dedicated HiperSockets *
  *                 device with QEBSM in a virtual machine.      *
  ****************************************************************
  * PROBLEM DESCRIPTION:                                         *
  ****************************************************************
  * RECOMMENDATION: APPLY PTF                                    *
  ****************************************************************
  The performance impact is caused by z/VM's processing of a QEBSM
  SIGA Instruction.  If z/VM handles the instruction, it may not
  reflect the short busy condition correctly to a guest.  This
  could occur when QIOASSIST is either turned off for a QEBSM
  Eligible HiperSockets device or if HiperSockets firmware taps
  z/VM on the shoulder to complete a SIGA when the assist is on.
  Not reflecting a short busy correctly will cause Linux's QDIO
  queue processing to get out of SYNC with the hardware, resulting
  in a drastic decrease in network performance from the point of
  failure on.
  
  Running with QIOASSIST ON, the chance of running into this
  problem is small.  With the assist on, HiperSockets firmware
  will process 99.9 percent of all QEBSM SIGA Instructions
  directly without z/VM involvement.  If a short busy condition is
  encountered while HiperSockets firmware is processing the
  instruction, it will be correctly presented to the guest.  Only
  the less than one percent of the cases where z/VM handles the
  SIGA Instruction itself is there a exposure for this problem.
  Given a short busy condition is not a typical event, the chances
  of z/VM running into the condition is very small with QIOASSIST
  ON.  The chance will increase considerably if QIOASSIST is
  turned off for the guest or system.  Turning off the assist will
  cause z/VM to execute all SIGA Instructions on behalf of the
  guest, thereby increasing the likelihood of running into the
  performance problem.
  

Problem conclusion

• Additional logic is added to z/VM's virtual SIGA Instruction
  logic to properly reflect a HiperSockets short busy condition
  for a guest exploiting a QDIO Enhanced Buffer State Management
  (QEBSM) network connection.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

  UM35032 UM35033 UM35034 UM35035

Modules/Macros

• HCPVQG
  

Fix information

• Fixed component name

  VM CP

• Fixed component ID

  568411202

Applicable component levels

• R540 PSY UM35032

     UP17/02/15 I 1000

• R620 PSY UM35033

     UP17/02/15 I 1000

• R630 PSY UM35034

     UP17/02/15 P 1701

• R640 PSY UM35035

     UP17/02/15 P 1701

Fix is available

• Select the PTF appropriate for your component level. You will be required to sign in. Distribution on physical media is not available in all countries.