Intelligent network release storage unlimited potential
dostor.com
Sun Ke
2016/12/27
On Dec. 9th, the China Storage Summit was held in Crowne Plaza Beijing, with the theme of "New Storage, New Normal, New Applications". China Storage Summit is the annual annual conference of the largest storage industry in Asia, which lasted for 12 years. It recorded the changes of the storage industry. The annual storage summit attracted the active participation of academia, industry and end-user representatives.
In memory of the summit "Flash technology" sub-forum keynote speech, director of marketing for network solutions Mai Siya Pacific Hui delivered a speech entitled " intelligent network storage release unlimited potential ," the keynote speech. The lecture focuses on how to free the performance of the storage over the network and release the performance of the flash to a higher level.
Director of Marketing, Solution Solutions, Asia Pacific
Hui said earlier conflicts are mainly concentrated in millisecond-level disk media latency , and network delay and software stack (200+ microseconds) are basically in microsecond level. However, the current media delay has been basically controlled at the microsecond level. This time you will find your bottleneck is not the principal contradiction in the media, but the network itself and the software stack.
The following is Zhang Hui's speech Record finishing:
Today I want to talk about the content of the network is how to release the performance of storage, and the performance of flash memory release higher.
We can now see that flash memory is growing faster and faster, the entire ecosystem will grow faster and faster, almost close to 100% growth rate. Of course, NVMe-related growth is also faster, in 2020, when 70% of the Server will use NVMe. Storage, almost 70,80% will be used NVMe. That is, less than NVMe scenes will be very small.
The earliest SAS, SATA, PCI NVMe. This technology do feel? Traditional disk, go down the SCSi protocol is no problem, but the media and technology in the development of the agreement (software layer) no greater progress, this time there will be reduced efficiency. There is a comparison, one is SAS, one is SATA, SATA-based PCIe performance will appear several times the gap, because NVMe the entire protocol layer to do the re-write, skip some of the more complex process.
 Storage is with the media changes, the development of science and technology, of course, the most important is also from changes in customer demand, forcing us to do better. To this end, we have two levels of coping methods. One is larger and the other is faster. Single medium is currently developing faster, including the future there will be 3D XPoint, and I personally believe that the future will be faster and faster Flash, getting better and better.
This problem is exposed to this single-node performance is getting better and better, then how to communicate between them? I give two examples, such as Beijing's traffic. I believe that both Beijing and three or four cities will be congested, widening the street is difficult to solve this problem. So how to solve this problem? Can not buy a car, just as impossible to not use our data, how to ease is also a problem. So I just mentioned the bandwidth, and then a protocol layer is the adjustment. In the network is relatively simple, we look at how Mellanox is done. NVMe appears in the future you will find that 10 trillion simply can not support, 100 G cases can only run 3-4. So, 25G just started, a 25G just enough to support a NVMe.
Just now we talked about the media, and now I want to talk about "super fusion." Hyper-fusion is a typical distributed, the communication between nodes will be great. Therefore, the technology of the network requirements will be more and more high, the delay and bandwidth are also high requirements.
Back to the technical look, we have to give priority to solve the main problems in the network. Early conflicts focused on millisecond disk media latency, and network latency and software stack (200 + microseconds) are basically in the microsecond level. However, the current media delay has been basically controlled at the microsecond level. At this time you will find your bottleneck is not the main contradiction of media, but the network itself and software stack. Mellanox has been able to control the software stack in the hundreds of nanoseconds. It can be said that the current TCP / IP methodology has affected the performance of the entire system.
What is RDMA? It is remote direct access, bypass the software stack, bypass the CPU, do not need CPU intervention, because the kernel will generate an interrupt, wait time appears. The RDMA is a direct remote access, CPU offload, the efficiency will improve many times, more than 200 microseconds software stack delay can be ignored.
This year there will be a new standard RoCE (RDMA over Converged Ethernet), is running on the Ethernet RDMA. Performance is slightly lower than running on InfiniBand, but relative to the previous line will have many times the upgrade, and lower cost. In June of this year there will be a new RoCE standard put forward, one of which is RoCE will support NVMe over Fabric. Our new products will support 200G.
In distributed storage, in order to ensure data security, will be backed up, but the profit margins will become low, you need to delete the code to improve the utilization rate. Erasure codes need to be recalculated, need CPU resources, the occupancy rate is sometimes very high. In our products, will delete the code to the flash card to do, do not take up CPU.
NVME Over Fabrics this well, the future space is also great. But in the traditional design, it also needs to consume CPU resources. Therefore, we hope that the CPU is only used for computing, such as network computing, NVME Over Fabrics these workloads, can be done to the flash card.
Just talk about some product and technology trends, and now I would like to introduce our products.
Today I want you to remember a new unit, that is, nanoseconds. Our latency for this machine is around 300 nanoseconds, which is more than 10 times worse than several subtle Ethernet products on the market, which is low latency. Bandwidth, the current is 100G, next year we will have 200G of the product launch. In addition, the packet loss is also affecting the storage efficiency of the more important issues, Mellanox can now do 0 packet loss. Of course, there are power issues, we also provide a very low power consumption.
Now I would like to talk about the issue of flash memory cards. I found a lot of technology in there. The first is the bandwidth problem, 100G bandwidth, delay 0.7 microseconds, or 700 nanoseconds, where our card latency is only 200 nanoseconds, the remaining delay generated by the software stack. Starting with ConnectX 4, we can support NVMe Over Fabrics. ConnectX 5 to this generation, began to support the NVMe uninstall, will release the CPU resources, so that your CPU specifically to do replication, snapshot, delete, disaster recovery, in order to achieve low latency, high efficiency. At the same time, ConnectX 5 also integrates E switch, based on this card can create a small storage system.
To the BlueField stage, it integrates ConnectX 5 all the features, 100G chips, front and rear ports, support for PCIe Gen3, PCIe Gen4, along with an ARM. It can be said that there is this chip, and a motherboard, you can design your entire system. BlueField can help enterprises to establish an all-flash system, from the front-end connection to the host, the back-end to the disk, and then to the calculation.
At present our latest product is ConnectX-6, supports PCIe Gen4, such as high-performance computing applications such as scenarios.
Here I would like to introduce something to the floor. The first Huawei, Huawei Mellanox based on this product, we do 10 million IOPS; The second is the Huayun Internet (FusionStack), single node to do 100 million IOPS; Here is the Memblaze, the same one million IOPS, bandwidth is 10GB / s The last one is Dell EMC's Isilon, is also used in our network. In the network based on Mellanox inside, can maximize the performance of the storage. We can see that all of our products are basically a single node million from IPOS.
I have a point in this, if you want to play to the extreme performance, then we must not let the network as your bottleneck. But to use the network, reduce your consumption, reduce your performance dependency, so that storage systems run faster.
thank you all.
This article copyright is stored online, not reproduced without permission. The article only represents the author's view, if you have different views, welcome to add storage micro-message public on-line (micro signal: doitmedia) to communicate. X, y |
