2. Objectives
Generate well-balanced
solutions without
compromising on availability,
performance, capacity and cost
Derive configurations from
empirical data during the
testing.
Provide reference architectures
to prime opportunities and
accelerate the sales cycle
3. Proof of Concept - Objective
Usable Capacity XS - 100TB S - 500TB M- 1PB L - 2PB
IOPS-Optimized
3x replicated data,
tiered over EC X X X X
Throughput
Optimized*
3x replicated data X X X X
Capacity Optimized*
6+2 EC protected data
X X X X
ScaleWorkload
The Ceph Performance Bingo Card
4. 36-Bay Server Flexibility
Framework to support multiple
configurations in the same box
Dual Socket Motherboard
Enabling Single and Dual CPU
Configurations
Dual Backplanes
x8 lane SAS Connectivity
Dual IT mode controllers
2x Available PCI slots for NVMe
Improving drive to SSD ratio
2x Dual port 10G cards
Separate Cluster and Client Networks
7. Write Journals for OSDs
can be stored on SSD
Media for optimal OSD
write performance.
The Ratio of SSD to HDD
allows Ceph to be tuned for
use in a broad spectrum of
applications
X9 Ceph Drive Configuration
5x OSDs 5x OSDs
SSD
SSD
OSD Node
5:1 Ratio
X9 generation
8. Write Journals for OSDs
can be stored on SSD
Media for optimal OSD
write performance.
The Ratio of SSD to HDD
allows Ceph to be tuned for
use in a broad spectrum of
applications
For 12-36 Bay OSD Nodes
the use of PCI-E flash /
NVMe instead of SAS
based SSD increases
available I/O bandwidth to
the media, enabling greater
HDD:SSD ratios to be used
X10 Ceph Drive Configuration
12 Bay Node (12+1)
12:1 Ratio
NVMe
NVMe
18:1 Ratio
NVMe
36 Bay Node (36+2)
X10 generation
13. Red Hat Ceph Reference Architecture
http://www.redhat.com/en/resources/red-hat-ceph-storage-clusters-supermicro-storage-servers
14. Proof of Concept - Objective
Usable Capacity XS - 100TB S - 500TB M- 1PB L - 2PB
IOPS-Optimized
3x replicated data,
tiered over EC X X X X
Throughput
Optimized*
3x replicated data X X X X
Capacity Optimized*
6+2 EC protected data
X X X X
ScaleWorkload
The Ceph Performance Bingo Card
15. Ceph Workload Positioning
Usable Capacity XS - 100TB S - 500TB M- 1PB L - 2PB
IOPS-Optimized
3x replicated data,
tiered over EC
Planned for Winter 2015
Throughput Optimized*
3x replicated data 6 node
• Read: 5,300 MB/s
• Write: 1,400 MB/s
32 node
• Read: 28,000 MB/s
• Write: 9,500 MB/s
63 node
• Read: 55,000 MB/s
• Write: 19,000 MB/s
125 node
• Read: 110,000 MB/s
• Write: 37,000 MB/s
Capacity Optimized*
6+2 EC protected data 10 node
• Read: 8,000 MB/s
• Write: 2,000 MB/s
8 node
• Read: 7,000 MB/s
• Write: 3,400 MB/s
13 node
• Read: 11,000 MB/s
• Write: 5,000 MB/s
* Projected performance based on node performance in 10 node cluster
(12 x 4TB + 1NVMe) (12 x 4TB + 1NVMe) (12 x 4TB + 1NVMe) (12 x 4TB + 1NVMe)
(12 x 6TB) (36 x 6TB) (72 x 6TB)
18. X10 and X9 Platforms for Ceph
X10 Models CPU/Mem
Drive
Config X9 Models CPU/Mem
Drive
Config
SSG-6018R-MON2
Dual Intel Xeon
E5-2630 v3
64GB
Ix 800GB PCI-flash /
NVMe
SSG-6017R-MON1
Dual Intel Xeon
E5-2630 v2
64GB
NA
SSG-F618H-OSD288P
Single Intel Xeon
E5-2620 v3
64GB
(12+1)
12x 6TB HDD + 1x
NVMe
NA NA NA
SSG-6028R-OSD072
Single Intel Xeon
E5-2620 v3
64GB
(12+0)
12x 6TB HDD
NA NA NA
SSG-6028R-OSD072P
Single Intel Xeon
E5-2620 v3
64GB
(12+1)
12x 6TB HDD + 1x
NVMe
SSG-6027R-OSD040H
Single Intel Xeon
E5-2630 v2
64GB
(10+2)
10x 4TB HDD +
2x SSD
SSG-6048R-OSD216
Dual Intel Xeon
E5-2630 v3
128GB
(36+0)
36x 6TB HDD
NA NA NA
SSG-6048R-OSD216P
Dual Intel Xeon
E5-2630 v3
128GB
(36+2)
36x 6TB HDD + 2x
NVMe
SSG-6047R-OSD120H
Dual Intel Xeon
E5-2630 v2
128GB
(30+6)
30x 4TB HDD +
6x SSD
SSG-6048R-OSD432
Dual Intel Xeon
E5-2690 v3
256GB
(72+0)
72x 6TB HDD
NA NA NA
SSG-6048R-OSD360P
Dual Intel Xeon
E5-2690 v3
256GB
(60+12)
60x 6TB HDD + 12x
SSD
SSG-6047R-OSD320H
Dual Intel Xeon
E5-2670 v2
(E5-2697
recommended)
128GB
(60+12)
60x 4TB HDD +
12x SSD
19. Conclusion
Single CPU 12+1 config seemed to be optimal for
throughput-optimized clusters.
SSD write journaling helps greatly. PCIe flash allows
greater HDD ratios for both throughput (replication)
and capacity (EC) optimized configs.
A lot more work is still needed
Identify IO optimized reference architecture
Study all flash and tiered performance
Understand OSD performance and implications of
non-uniform memory access (NUMA) systems
Proc-core to OSD ratios with/without hyper-threading