Over the past years, cloud storage highly increased its performance by establishing NVMe-SSD as a standard. Dropping conventional serial (or SATA) has allowed stepping up to next level of performance — increased bandwidth, much more IOPS and improved efficiency under storage related workloads. The growing need of lower latency drives and boosted throughput for users with disk intensive database, machine learning or big data workloads has been observed as bottleneck within the market. OVHCloud proposes a solution with their new I1 series of High IOPS virtual machines.
They explained in an article, how they design their architecture to be the most efficient as possible. Though not exclusive to OVHCloud via patent, the following feature set gives a tremendous leg up over rivals:
- NVMe cards sized at 1.9TB
- Single-tenant volume access
- Local interface avoiding network overhead of block storage
- PCI passthrough, circumventing a layer of virtualization for direct access, improved driver support and dramatic increases in performance potential
This series is actually available in 3 flavors, ranging from 8 to 32 CPUs with equivalent of CPU/RAM/Storage ratios:
Name | CPU | RAM | Root volume | Extra volume(s) | Storage size |
---|
I1-45 | 8 | 45 | 50GB Root local SSD | 1x 1.9TB I1 NVMe SSD | 1.9TB |
I1-90 | 16 | 90 | 50GB Root local SSD | 2x 1.9TB I1 NVMe SSD | 3.8TB |
I1-180 | 32 | 180 | 50GB Root local SSD | 4x 1.9TB I1 NVMe SSD | 7.6TB |
Market analysis
Cloud Spectator was commissioned by OVHcloud to perform a market analysis of rival offerings
from some large competitors with similar NVMe storage implementations. The Cloud Spectator
analysis was based on three metrics: raw performance (random operations), price and value (ratio
of price and performance). OVHcloud claims they are trying to provide the best performance in
the cloud sphere, avoid cost prohibitive offerings and make new technology widely available to
customers. Those claims were scrutinized with a close eye, redundant sampling and benchmarks.
The evaluation found not only fast drives, reasonable prices and remarkable value compared to
the field of rival offerings, but dramatically so. Do not take these simple words for it, the
benchmarks results are straight ahead:
Provider | Name | Type | CPU | RAM | Storage |
---|
OVHCloud | I1-45 | High IO | 8 | 45 | 1x 1,9TB |
OVHCloud | I1-90 | High IO | 16 | 90 | 2x 1,9TB |
Microsoft Azure | Standard_L8s_v2 | High IO | 8 | 64 | 1x 1,7TB |
T-Systems OTC | h2.3xlarge.10 | HPC | 16 | 128 | 1x 3,2TB |
Oracle Cloud | VM.DenseIO1.4 | High IO | 8 | 60 | 1x 3,2TB |
Amazon Web Services | i3.2xlarge | High IO | 8 | 64 | 1x 1,9TB |
All evaluated machines have at least one NVMe-SSD and minimum of 8 virtual CPUs. Comparable machines of the I1-45 consisted of those with single 1.9TB devices while VMs with minimum volume size of 3.2TB were chosen for comparison with the OVHCloud’s I1-90 2×1.9 (3.8TB).
Performance
It was surprising to see OVHcloud NVMe performance compared to the top-tier offerings from massive players in the cloud industry. Although no provider had poor performance, the results found the Oracle, Azure and AWS could not keep pace with OVHCloud implementations. OVHcloud simply outperformed everything else and at delivers the fastest NVMe storage we’ve observed in the cloud market. Each OVHCloud 1.9TB pushed ~600,000 read IOPS and 450,000 write IOPS. Substantially more were squeezed from the 2 drives configuration of the I1- 90 with MDADM software striping, even overshadowing one dedicated High-Compute (HPC) VM.
{"configuration":{"chart":{"type":"bar","polar":false,"options3d":{},"height":null,"inverted":false},"credits":{"enabled":false},"title":{"text":"IOPS performance"},"colorAxis":null,"subtitle":{"text":""},"xAxis":{
"title":{"text":""},"categories":["Read", "Write"]},"yAxis":{"title":{"text":"in IOPS"},"categories":null,"plotLines":null,"plotBands":null,"floor":0,"gridLineInterpolation":null},"zAxis":{"title":{"text":"IOPS"}},"plotOptions":{"series":{"dataLabels":{"enabled":false,"format":null}},"column":{"stacking":""},"area":{"stacking":""}},"rangeSelector":{"enabled":false},"legend":{"enabled":true},"series":[
{"id": "i1-45", "name": "I1-45", "color": "#484848", "data": [593066, 459464]},
{"id": "i1-90", "name": "I1-90", "color": "#686868", "data": [1153783, 509589]},
{"id": "h2.3xlarge.10", "name": "h2.3xlarge.10", "color": "#d1006c", "data": [773412, 484706]},
{"id": "i3.2xlarge", "name": "i3.2xlarge", "color": "#f7981d", "data": [413307, 182164]},
{"id": "Standard_L8s_v2", "name": "Standard L8s v2", "color": "#55b3ff", "data": [456240, 245376]},
{"id": "VM.DenseIO1.4", "name": "VM.DenseIO1.4", "color": "#FF0000", "data": [304820, 280947]}
],"tooltip":{"enabled":false}},"hc_type":"chart","id":"140417476574680"}
Results given here are collected with our standard benchmark methodology, we use FIO in random mode, with 4KB blocks. The volume is tested directly without filesystem and with flags indicating a minimum cache/buffer implication. Volumes also have been pre-warmed entirely in write/read.
Price
Price often determines exclusivity of a product, some may being more expensive (often a
luxury), some more economical (but can they handle the task given?) — a healthy balance of both is desirable for most cloud consumers, now for those who want everything, there is always High-Performance Compute, deep pockets may be required. For the rest of us, let’s take a look at value of the evaluated machines.
Provider | Name | Hourly | Monthly (730 hours) |
---|
OVHCloud | I1-45 | $0,44 | $324,12 |
OVHCloud | I1-90 | $0,89 | $648,97 |
Microsoft Azure | Standard_L8s_v2 | $0,74 | $540,2 |
T-Systems OTC | h2.3xlarge.10 | $2,22 | $1 620,60 |
Oracle Cloud | VM.DenseIO1.4 | $0,51 | $372,30 |
Amazon Web Services | i3.2xlarge | $0,69 | $503,70 |
Based on hourly pricing OVHcloud leads this ranking, it may not seem a frank length on hourly basis but shifted to monthly billing the gap is widening. The price affect on value is clearly observed in the next section.
Remarks: No discount have been considered here, most providers have discount offers depending on the engagement. For instance on AWS you could have around 30% for a yearly engagement, on OVHcloud you could have also around 30% for a monthly engagement.
Price/performance
Finally, the price-performance ratio will give us an apples-to-apples comparison on machine value per USD spent. Not as large a surprise as seen in the raw results, OVHcloud VMs costs substantially less while putting out superior raw performance numbers to rival matches. In short, OVHCloud hardware gave both the best economy and performance:
{"configuration":{"chart":{"type":"bar","polar":false,"options3d":{},"height":null,"inverted":false},"credits":{"enabled":false},"title":{"text":"Price/performance score"},"colorAxis":null,"subtitle":{"text":""},"xAxis":{
"title":{"text":""},"categories":["Read", "Write"]},"yAxis":{"title":{"text":"in IOPS"},"categories":null,"plotLines":null,"plotBands":null,"floor":0,"gridLineInterpolation":null},"zAxis":{"title":{"text":"IOPS"}},"plotOptions":{"series":{"dataLabels":{"enabled":false,"format":null}},"column":{"stacking":""},"area":{"stacking":""}},"rangeSelector":{"enabled":false},"legend":{"enabled":true},"series":[
{"id": "i1-45", "name": "I1-45", "color": "#484848", "data": [1830, 1418]},
{"id": "i1-90", "name": "I1-90", "color": "#686868", "data": [1778, 785]},
{"id": "h2.3xlarge.10", "name": "h2.3xlarge.10", "color": "#d1006c", "data": [477, 299]},
{"id": "i3.2xlarge", "name": "i3.2xlarge", "color": "#f7981d", "data": [821, 362]},
{"id": "Standard_L8s_v2", "name": "Standard L8s v2", "color": "#55b3ff", "data": [845, 454]},
{"id": "VM.DenseIO1.4", "name": "VM.DenseIO1.4", "color": "#FF0000", "data": [819, 755]}
],"tooltip":{"enabled":false}},"hc_type":"chart","id":"140417476574685"}
Appendix
This article has been written by Cloud Spectator, an objective, third party market research and benchmarking firm specializing in analysis of the cloud market and its offerings. Further information regarding this study, including detailed test methodology and detailed analyses can be located in the following the
PDF report.