mariadb on ibm power8

4

Upload: kim-kozlik

Post on 08-Feb-2017

36 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: MariaDB on IBM POWER8

MariaDB - 64% faster on POWER8 than on Intel

MariaDB 1 0.1 can process 1 .2 mil l ion queries per second

The pressure on costs is endless and at Quru, we bel ieve it is system ‘consol idation’ that

potential ly promises major reductions in data centre costs (space + power + cool ing)

together with lower service contract costs. We have recently been focused on consol idating

servers and adding extra compute capacity in the same footprint as a way to enable lower

costs and increase appl ication performance.

However, consol idation requires a platform that can rapidly and easi ly both scale-up and

scale-out. This technical Insight white paper covers benchmark testing we have undertaken

of MariaDB on IBM POWER8 - a combination that promises a step change in consol idating

systems and the opportunity to significantly enhance appl ication performance. Our work in

tuning workloads has demonstrated that MariaDB on POWER8 can lead to lower costs and

increased datacentre efficiency.

+44 (0) 20 71 60 2882@qurul inkedin.com/company/[email protected]

MariaDB on IBM POWER8Benchmark Performance Testing

In partnership with:

Page 2: MariaDB on IBM POWER8

QURU INSIGHTS 2QURU INSIGHTS

The IBM PowerLinux platform, based on the POWER8 series of processors, provides unparal leled levels of

performance and scalabi l i ty with features not avai lable in other mainstream platforms. Of these perhaps the most

significant is that the POWER8 processors include up to 8 hardware threads per CPU core, hence a humble 2U, 2-

socket system can del iver up to 1 60 threads. Today, no other vendor can offer this capabi l i ty.

MariaDB 1 0.1 contains many new features that enable high performance – in particular scalabi l i ty on massively

multithreaded hardware. So, MariaDB have already publ ished benchmark results showing that MariaDB 1 0.1 running

on POWER8 systems is capable of 1 .2M queries per second when al l 1 60 threads are employed.

The next stage is to identify how the above

performance of MariaDB 1 0.1 on POWER8 compares

to the Intel based systems in the market. today.

There are many ways to benchmark a system, but at

present a 2-socket platform from Intel providing 1 60

threads is not avai lable. A typical system at a simi lar

price point to the IBM Power S822L benchmarked in

our tests may wel l use two Haswel l v3 architecture

Xeon E5 processors, featuring 1 0 cores each, with two

hardware threads per core, i .e. a total of 40 threads.

Our test system was a SuperMicro server and had 40

hardware threads, 1 28GB RAM, and SSD backed

storage, and employed the E5-2650 v3 chip, running at

2.30GHz (with turbo boost to 3.0GHz) .

To perform a directly comparable test, we configured a

PowerKVM virtual machine on our S822L featuring 40

threads (2 sockets, 1 0 cores per sockets, 2 threads per

core – directly comparable to the Intel platform) and

1 28GB RAM. To remove variabi l i ty of test results based

on disk performance variations, the test database was

constructed in memory and run from there on both

platforms. Whi lst this would not be a real world setup, it

removes from the equation any possible storage

bottlenecks - it is fair to assume that a performance

database would be stored on a high performance

storage array where I/O bottlenecks are less of an issue.

MariaDB 1 0.1 was then configured identical ly on each

system and the sysbench 0.5 benchmark run using

identical settings on both systems to produce

comparable benchmarks. The results are quite notable.

POWER8 & MariaDB 1 0.1 vs. Intel benchmark test

We can see that with a read-only OLTP test, the S822L system is approximately 64% faster thanthe Intel system, in spite of having exactly the same number of threads and the same amount of

memory.

Page 3: MariaDB on IBM POWER8

QURU INSIGHTS 3

This is not the end of the story. The POWER8

platform development toolchain, provides access to a

suite of tools not avai lable on other platforms, and

which enable powerful tuning and optimization

techniques to be performed on existing software.

Taking advantage of this capabi l i ty, we then undertook

a tuning and optimization process specifical ly on the

MariaDB test workload we had already used and

achieved results 64% faster than on Intel .

This second testing exercise - which we have not seen

anywhere else in the market at this time - produced a

dramatic performance improvement over the original

bui ld of MariaDB we tested, achieving benchmarks of at

least 1 5% faster than seen on the un-tuned software.

Adding this to the high performance of the POWER8

platform results in a gain of 93% over the reference

Intel platform, even though both systems have the same

number of hardware threads enabled.

Unique Testing Exercise

Going back to the first testing we undertook, our S822L

had only 40 of its 1 60 threads in use, hence it would be

possible to either scale up or scale out the MariaDB

virtual machine; after al l we already know that at 1 60

threads, more than 1 000,000 QPS is achievable under

the same test conditions as performed here.

With the improved performance of POWER8 over Intel

Haswel l v3 for this workload, it may be that the

configuration is entirely sufficient. I f so, there is ample

headroom for additional VMs on the S822L, for

example to run 4 equivalent database VMs.

Alternatively, if appl ications are not performing wel l ,

more threads/CPU can be al located to enhance their

performance, scal ing MariaDB 1 0.1 from 64% faster than

Intel using 40 threads to a massive 1 25% faster when al l

1 60 threads are al located to MariaDB.

Hence it is l ikely that many use cases wi l l attain the

performance they require whi lst using only a fraction of

the hardware resources avai lable on the POWER8

platform. This releases capacity to run more VMs and

appl ications alongside it and means that consol idation

through MariaDB on POWER8 real ly can lead to lower

costs and increased efficiency in the datacentre.

In contrast, a 40-thread Intel based system would be

almost ful ly uti l ized by the above configuration and

hence the only way to scale up, or scale out, would be

to purchase additional hardware. This would obviously

involve increased capital and operational expenditure,

and require more datacenter space, cool ing, and power.

MariaDB is proven to run at 1M QPS on POWER8 when all 160 threads are dedicated to it -however the Quru optimised version of MariaDB will run at 1.2M QPS

Page 4: MariaDB on IBM POWER8

QURU INSIGHTS 4

IBM’s Power System is a mature, highly scalable and

high performance RISC architecture featuring multi-

core, multi-terabyte memory servers combined with

open software architectures.

MariaDB is the trusted adviser for the world’s leading

MariaDB and MySQL users with a rich heritage of

database expertise. MariaDB Enterprise forms part of

the most robust ecosystems in open source software

with strong endorsement from leading technology

vendors such as Red Hat, IBM, SuSE, Intel and Google.

MariaDB is an integral part of IBM’s TurboLAMP stack,

a ful l infrastructure stack optimized for POWER8.

MariaDB has been ported to Linux on Power, but there

is more to it than just recompi l ing the code. MariaDB is

bui lt using IBM Advance ToolChain on POWER8 to

ensure maximum stabi l i ty and performance. In addition,

MariaDB has added code changes to optimize for use

with the POWER8 architecture, increase its stabi l i ty and

to take ful l benefit of POWER8 features.

Benchmarks performed by MariaDB Corporation and

IBM show MariaDB del ivering nearly four times the

single-core performance on POWER8 vs. equivalent

x86-based servers. Hence, MariaDB Enterprise is the

ideal subscription to increase the performance of single

servers on POWER8 in a scale-out architecture.

MariaDB was bui lt from source using the IBM Advance

Toolchain version 8.0. Binaries were bui ld unstripped.

IBM's FDPR tool (Feedback Directed Program

Restructuring) was then run against the mysqld binary

produced to optimize the binary against the sysbench

workload, and the benchmarks were repeated.

For more detai ls on benchmark configuration and setup

as used in this whitepaper, please visit the fol lowing l ink:

http://quru.com/whitepapers/madiadbonpower8/

We are confident that the IBM POWER8 platform

offers significant opportunities for consol idating

workloads or achieving higher performance on existing

workloads, or any combination of the two.

Tuning and optimization such as this is highly workload

dependent, and each workload wi l l have to be tuned

differently to achieve the best results. Previous work by

Quru has proved that tuning can vastly improve the

performance of workloads on the IBM POWER

platform and we are currently working with IBM on

tuning these workloads to achieve even higher

performance levels without any changes to hardware

specification.

Quru is very wel l placed to do this, with extensive

experience in tuning and optimizing software for the

IBM PowerLinux platform, and can help your business

achieve even better results from the IBM POWER8

platform.

For more information on how your organisation can

improve the performance of MariaDB on IBM

POWER8, cal l Robin Porter on +44 (0) 20 71 60 2882.

Quru - experts in performance gains via tuning

About Quru

Quru is a market leader in the technicaldevelopment, deployment and support of Linuxand open source solutions that help organisationsto reduce costs and increase operational agi l i tyand capabi l i ty. We have also developed multipleaward-winning software solutions ranging frommobi le phone apps to global enterprise systems.

+44 (0) 20 71 60 2882@qurul inkedin.com/company/[email protected]

Market leading technology - IBM POWER

MariaDB optimized for POWER8