2009-01-28 DOI NBC Red Hat on System z Performance Considerations

Red Hat on System z
Performance Considerations
Updated 27-JAN-2009

Agenda & IntroductionAgenda & Introduction
● Networking
● Storage
● Memory Management
Much of this presentation is a collection of IBM
reports, SHARE presentations, and Red Hat data.
This presentation puts it in one place.

Agenda & IntroductionAgenda & Introduction
Shawn Wells <swells@redhat.com>
Global System z Alliance Manager
(+1) 443 534 0130
● Based in Washington, D.C.
● Global responsibility for Red Hat's System z activities

Who's seen this before?Who's seen this before?

Network PerformanceNetwork Performance
ConsiderationsConsiderations

The EnvironmentThe Environment
● HiperSockers
– Four Linux LPARs (2 IFLs ea.) to one z/OS Server
LPAR (4 & 8 Cps)
– z/OS V1R2
– TCP buffer sizes increased from 65335B to
131071B, allowing for more unacknowledged data
– L3 traffic only
– IPv4, z990, Large Send (requires OSA, card-to-card,
and IPv4)

● Throughput & Response time charts
– Transactions per second for interactive workloads in
milliseconds
– Megabytes per second for streams-type workloads
● 1MB = 1,048,576 bytes/sec
● Throughput & CPU-Milliseconds Charts
– # of milliseconds of processing per transaction
– Streams-type workloads, # milliseconds of
processing per megabyte

WORKLOAD DESCRIPTION
cra CRR 200B/15000B (A 200-byte Request-message from the client to the
server followed by a 15000-byte Response-message from the server to
the client. A TCP-connection is established for each RR.)
rra RR 200B/1000B (A 200-byte Request-message from the client to the
server followed by a 1000-byte Response-message from the server to the
client. All RRs operate under that same TCP-connection.)
rrb RR 200B/15000B
rr1 RR 1B/1B
rr64k RR 64KB/64KB
sor Streams Outbound (20B from the client to the server, followed by 20MB
from the server to the client)
sir Streams Inbound (20MB/20B)
mix Mixe of sor and sir streams

OSA Response TimeOSA Response Time
200 byte request, 15000 byte response, same tcp session200 byte request, 15000 byte response, same tcp session

Transactions per Second vs CPU TimeTransactions per Second vs CPU Time

Streaming Data @ 57344B/secStreaming Data @ 57344B/sec

Storage PerformanceStorage Performance

● Configuration for 4Gbps disk I/O measurements

Disk I/O Performance with 4Gbps links –Disk I/O Performance with 4Gbps links –
FICON vs FCPFICON vs FCP
● Throughput for sequential write is simular
● FCP throughput for random I/O is 40% higher

Disk I/O Performance with 4Gbps links –Disk I/O Performance with 4Gbps links –
FICON vs FCP / direct I/OFICON vs FCP / direct I/O
● Bypassing RHEL page cache improves throughput for FCP by 2x,
FICON up to 1.6x
● Read operations much faster on FCP

FCP/SCSI single path vs multipath LVMFCP/SCSI single path vs multipath LVM
● Use failover instead of multibus

StripingStriping
● Results with SCSI disks
– LVM striped logical volumes are fast, but LVM takes
up % more CPU
● Results with ECKD disks
– Fairly the same outcome as with SCSI disks.
Performance differences with storage pool striping
were small
– The load balancing over a FICON path group
performs better than dm-multipath (RHEL
multipathing)

Crypto PerformanceCrypto Performance

Crypto Express2 Accelerator (CEX2A)Crypto Express2 Accelerator (CEX2A)
SSL handshakesSSL handshakes
● The number of handshakes is up to 4x higher with h/w assist
● In 32 connections case, we save 50% of CPU

Crypto Performance – WebSEAL SSLCrypto Performance – WebSEAL SSL
AccessAccess
● Improvement by h/w support

Memory ManagementMemory Management
Performance ConsiderationsPerformance Considerations

CMM1 & CMMACMM1 & CMMA
● Cooperative Memory Management (CMM1)
– Ballooning technique
– When z/VM detects memory constraints, it tells
RHEL guests to release page frames (done by
issuing a DIAG X'10')
– When memory is freed, it alerts requesting RHEL
guest

CMM1 & CMMACMM1 & CMMA
● Collaborative Memory Management Assist
(CMMA)
– Page status technique
● Stable (S)
page has essential content
● Unused (U)
no useful content and any access to the page will
cause an addressing exception
● Volatile (V
page has useful content. CP can discard the page
anytime.

CMM & CMMA PerformanceCMM & CMMA Performance

Open Discussion / Q&AOpen Discussion / Q&A

2009-01-28 DOI NBC Red Hat on System z Performance Considerations

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to 2009-01-28 DOI NBC Red Hat on System z Performance Considerations (20)

More from Shawn Wells (20)

Recently uploaded (20)

2009-01-28 DOI NBC Red Hat on System z Performance Considerations