Scylla Summit 2017: Intel Optane SSDs as the New Accelerator in Your Data Center

PRESENTATION TITLE ON ONE LINE
AND ON TWO LINES
First and last name
Position, company
Intel® Optane™ SSDs and Scylla
Providing the Speed of an In-memory
Database with Persistency
Tomer Sandler and Frank Ober

AND ON TWO LINES
First and last name
Position, company
Tomer Sandler
Solution Architect @ ScyllaDB
2
Data Center Solution Architect @ Intel®
Frank Ober

AND ON TWO LINES
First and last name
Position, company
Agenda
▪ Introduction
▪ Intel® Optane™ SSD DC P4800X
▪ Scylla as an In-Memory Like Solution
▪ How We Knew Optane™ is Going to “Rock”
▪ Setup and Workloads
▪ Results
▪ TCO: Enterprise SSD vs. Intel® Optane™
▪ Summary
3

AND ON TWO LINES
First and last name
Position, company
Introduction
The Challenge
Providing a solution with the performance of an in-memory like
database without compromises on throughput, latency, and data
persistence.
4

AND ON TWO LINES
First and last name
Position, company
Introduction
The Challenge
Providing a solution with the performance of an in-memory like
database without compromises on throughput, latency, and data
persistence.
How...
Using Scylla and Intel® Optane™ SSD DC P4800X to resolve cold-cache
and data persistence challenges.
5

AND ON TWO LINES
First and last name
Position, company
Intel® Optane™ SSD DC
P4800X

AND ON TWO LINES
First and last name
Position, company
7

AND ON TWO LINES
First and last name
Position, company
8

AND ON TWO LINES
First and last name
Position, company
9

AND ON TWO LINES
First and last name
Position, company
10

AND ON TWO LINES
First and last name
Position, company
11

AND ON TWO LINES
First and last name
Position, company
12

AND ON TWO LINES
First and last name
Position, company
Scylla as an
In-Memory Like
Solution

AND ON TWO LINES
First and last name
Position, company
Scylla as an In-Memory Like Solution
▪ In-Memory Database Requirements
o Sub-millisecond response time
o High throughput
o Support large number of clients concurrently
14

AND ON TWO LINES
First and last name
Position, company
▪ In-Memory Database Requirements
o Sub-millisecond response time
o High throughput
o Support large number of clients concurrently
▪ In-Memory Database Challenges
o Cold cache and long warmup times
o Persistency and high availability
o Scalability
o Simplistic data models
15

AND ON TWO LINES
First and last name
Position, company
▪ Scylla provides
o Persistent data storage
o High throughput, low latency data access
o Rich data model capabilities
▪ Scylla scales (and scales...)
▪ Scylla needs VERY fast storage media to pair with
▪ Ease fetching and storing information latency
16

AND ON TWO LINES
First and last name
Position, company
How We Knew
Optane™ is Going
to “Rock”

AND ON TWO LINES
First and last name
Position, company
How We Knew Optane™ is Going to “Rock”
▪ We used Diskplorer to measure the drives capabilities
o Small wrapper around fio that is used to graph the relationship between
concurrency (I/O depth), throughput, and IOps
18

AND ON TWO LINES
First and last name
Position, company
o Concurrency is the number of parallel operations that a disk or array can
sustain. With increasing concurrency, the latency increases and we observe
diminishing IOps increases beyond an optimal point
19

AND ON TWO LINES
First and last name
Position, company
o Concurrency is the number of parallel operations that a disk or array can
sustain. With increasing concurrency, the latency increases and we observe
diminishing IOps increases beyond an optimal point
RandRead test with a 4K buffer:
● Optimal concurrency is ~24
● Throughput: 1.0M IOps
● Latency: 18µs
20

AND ON TWO LINES
First and last name
Position, company
Setup and Workloads

AND ON TWO LINES
First and last name
Position, company
Setup and Workloads
▪ 3 Scylla v2.0 RC servers: 2 x 14 Core CPUs, 128GB DRAM, 2 x Intel®
Optane™ SSD DC P4800X
o CPU: Intel® Xeon® CPU E5-2690 v4 @ 2.60GHz
o Storage: RAID-0 on top of 2 Optane™ drives – total of 750GB per server
o Network: 2 bonded 10Gb Intel® x540 NICs. Bonding type: layer3+4
▪ 3 Client servers: 2 x 14 Core CPUs, 128GB DRAM, using the
cassandra-stress tool with a user profile workload
▪ Set the # of IO queues equal to the # of shards
o /etc/scylla.d/io.conf: SEASTAR_IO="--num-io-queues=54
--max-io-requests=432"
22

AND ON TWO LINES
First and last name
Position, company
Setup and Workloads
▪ Cassandra-stress: User defined mode that allows running
performance tests on custom data models, using yaml files for
configuration
▪ Simple K/V schema used to populate ~50% of the storage capacity
▪ Utilizing all of the server’s RAM (128GB), replication factor set to 3
(RF=3), and the consistency level is set to one (CL=ONE)
▪ Tested 1 / 5 / 10 KByte payloads
o Challenge the default 512B sector size
o Max. IOps for each payload, at very low latency for reads
23

AND ON TWO LINES
First and last name
Position, company
Setup and Workloads
▪ Two scenarios for read tests
o Large working set much larger than the RAM capacity. This scenario lowers the
probability of finding a read partition in Scylla’s cache
o Small working set that will create a higher probability of a partition being
cached in Scylla’s memory
▪ Latency measurements
o Cassandra stress client end-to-end latency results
o Scylla-server side latency results (using `nodetool tablehistograms` command)
24

AND ON TWO LINES
First and last name
Position, company
Results

AND ON TWO LINES
First and last name
Position, company
Latency Test Results
26
Payload Size Test Case (RF=3)
Total Requests
per Sec
Cassandra stress 95%
Latency (ms)
Scylla-server 95%
Latency (ms)
Disk Throughput per
Server (GBps)
Load per
Server
1 KB
key:64b
blob:1kb
Write
300M Partitions
(~50% disk space)
Avg: ~196K
Max: 220K
2.0
Avg: ~1.25
Max: 2.65
~65%
Read
Large Spread
(~75% from Disk)
198K 0.7 0.478
Avg: ~1.65
Max: 2.2
~32%
Read
Small Spread
(All in-Memory)
198K 0.4 0.023 None ~15%
5 KB
key:64b
blob:5kb
Write
75M Partitions
(~54% disk space)
Avg: ~166K
Max: 180K
2.8
Avg: ~2.75
Max: 4.2
~65%
Read
Large Spread
(75% from Disk)
168K 0.9 0.405
Avg: ~1.22
Max: 1.84
~36%
Read
Small Spread
(All in-Memory)
168K 0.5 0.0405 None ~18%

AND ON TWO LINES
First and last name
Position, company
Latency Test Results
27
Total Requests
per Sec
Latency (ms)
Scylla-server 95%
Latency (ms)
Disk Throughput per
Server (GBps)
Load per
Server
10 KB
key:64b
blob:10kb
Write
36M Partitions
(~50% disk space)
120K 2.45
Avg: ~3.7
Max: 4.5
~65%
Read
Large Spread 1
(75% from Disk)
120K 1.0 0.398
Avg: ~0.95
Max 1.72
~30%
Read
Large Spread 2
(75% from Disk)
166K 1.2 0.481
Avg: ~1.35
Max: 2.27
~40%
Read
Small Spread
(All in-Memory)
166K
(120K)
0.6
(0.5)
0.063
(0.051)
None ~22%

AND ON TWO LINES
First and last name
Position, company
Throughput Test Results
28
Total Requests
per Sec
Latency (ms)
Cassandra stress
threads per client
Disk Throughput per
Server (GBps)
Load per
Server
128B
key:64b
blob:128b
Write
600M Partitions
(~8% disk space)
Avg: ~1.95M
Max: 3.05M
7.3 520
Avg: ~0.55
Max: 1.12
~95%
Read 300M
Large Spread
(~50% from Disk)
Avg: ~976K
Max: 1.35M
2.5 120
Avg: ~2.3
Max: 4.29
~94%
Read 600M
Large Spread
(~60% from Disk)
Avg: ~771K
Max: 986K
2.95 120
Avg: ~3.35
Max: 4.53
~94%
Read
Small Spread
(All in-Memory)
Avg: ~2.19M
Max: 2.21M
2.6 300 None ~96%
▪ 128B payload with RF and CL = ONE
▪ 12 cassandra-stress instances (each instance populating a different range).
▪ Read large spread test ran twice, once on the full range (600M partitions) and once on half the
range (300M partitions)

AND ON TWO LINES
First and last name
Position, company
Throughput Test Results
29
Total Requests
per Sec
Latency (ms)
Cassandra stress
threads per client
Disk Throughput per
Server (GBps)
Load per
Server
128B
key:64b
blob:128b
Write
600M Partitions
(~8% disk space)
Avg: ~1.95M
Max: 3.05M
7.3 520
Avg: ~0.55
Max: 1.12
~95%
Read 300M
Large Spread
(~50% from Disk)
Avg: ~976K
Max: 1.35M
2.5 120
Avg: ~2.3
Max: 4.29
~94%
Read 600M
Large Spread
(~60% from Disk)
Avg: ~771K
Max: 986K
2.95 120
Avg: ~3.35
Max: 4.53
~94%
Read
Small Spread
(All in-Memory)
Avg: ~2.19M
Max: 2.21M
2.6 300 None ~96%
▪ 128B payload with RF and CL = ONE
▪ 12 cassandra-stress instances (each instance populating a different range)
▪ Read large spread test ran twice, once on the full range (600M partitions) and once on half the
range (300M partitions)

AND ON TWO LINES
First and last name
Position, company
TCO

AND ON TWO LINES
First and last name
Position, company
TCO: Enterprise SSD vs. Intel® Optane™
Intel® Optane™ provide great latency results, and is also more than
50% cheaper compared to DRAM or Enterprise SSD configurations
31

AND ON TWO LINES
First and last name
Position, company
TCO: Enterprise SSD vs. Intel® Optane™
Intel® Optane™ provide great latency results, and is also more than
50% cheaper compared to DRAM or Enterprise SSD configurations
32

AND ON TWO LINES
First and last name
Position, company
Summary

AND ON TWO LINES
First and last name
Position, company
What did we learn
▪ Scylla’s C++ per core scaling architecture and unique I/O scheduling can
fully utilize your infrastructure’s potential for running high-throughput
and low latency workloads
▪ Intel® Optane™ and Scylla achieve the performance of an all in-memory
database
▪ Intel® Optane™ and Scylla resolve the cold-cache and data persistence
challenge without compromising on throughput, latency and performance
▪ Data resides on nonvolatile storage
▪ Scylla server’s 95% write/read latency < 0.5msec at 165K requests per sec
▪ TCO: 50% cheaper than an all in-memory solution
34

AND ON TWO LINES
First and last name
Position, company
THANK YOU
Tomer@scylladb.com
Please stay in touch
Any questions?
Frank.Ober@intel.com
Check our blogs
- Intel Optane Review
- Intel Optane and Scylla

Scylla Summit 2017: Intel Optane SSDs as the New Accelerator in Your Data Center

Related slideshows

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (7)

Similar to Scylla Summit 2017: Intel Optane SSDs as the New Accelerator in Your Data Center

Similar to Scylla Summit 2017: Intel Optane SSDs as the New Accelerator in Your Data Center (20)

More from ScyllaDB

More from ScyllaDB (20)

Recently uploaded

Recently uploaded (20)

Scylla Summit 2017: Intel Optane SSDs as the New Accelerator in Your Data Center