KB Article #179015

cpu consumption higher on the receiver than the sender for SSL transfers

Problem

Moving from non-SSL to SSL transfer in between two LPAR


Hard crypto is no enabled on both LPARs


Same OS level and Transfer CFT version on both side, same parameters.


CPU consumption on the receiver is about 3 time when compared to a non-SSL transfer.


CPU consumption on the sender is only 1.4 time higher when compared to a non-SSL transfer.


System Hardware on the sender is a z13 while a z196 is used on the receiver side.

Resolution


The difference of performances in between the z13 architecture and the older z196 explain it all (since no hardcrypto used on both LPAR).


SSL is CPU hungry when no hardcrypto is available to help and the performance differences in the 2 systems architecture is already close to the difference observed by the customer.


A 3 times average more CPU with SSL vs non-SSL, it is just a price to pay for SSL for before z13 architectures.


FYI, main changes in between z13 and z196 below:


z13 improves further on top of the zEC12 design
Runs at a slightly lower maximum frequency of 5 GHz; with a much wider pipeline (2x) to handle more instructions per cycle for a net increase in overall instruction execution rate
Integrates L2 cache lookup into L1 instruction cache design to improve instruction L2 cache access latency
Supports simultaneous multi-threading (SMT) for 2 threads
Introduces Single-Instruction-Multiple-Data (SIMD) instructions for vector operations


------------

Z196 introduces the first generation out of order pipeline design
Runs at 5.2 GHz on the EC class machines
Introduces high-word architecture with operations on upper 32 bits of general purpose registers (GRs)
Adds more nondestructive arithmetic instructions
Adds conditional load and store instructions, for reducing potential branch wrong penalties