July 7, 2016 at 09:53

Quanta server overview: interesting solutions, rack-sized blade baskets, microserver hives and HPC farms


    A tray with disks leaves the case, and the server remains in the rack.

    On the server market, finding something new is a great success, because most of the evolution revolves around well-known technologies, just making them stronger, higher, faster and cheaper. When the Quanta servers were brought to us, I expected everything that was familiar, but with a different interface. However, the glands that were brought to CROC for testing could surprise me.



    The first surprise was thin “in the ass” SSD disks of servers, they are convenient to use for OS or data caching. The second - on one of the models, receivers for HDDs conveniently leaving the server. Well, and a bunch of different little things that speak of a really well-thought-out assembly, in which, it seems, not only the manufacturer’s engineering brain participated, but also people who know firsthand how to administer farms and data centers. As you review, I’ll show you some more examples.

    It is noteworthy that Quanta did not release clones of what is on the market. I'll start with a story about unusual models, then I will tell you more about the iron that arrived. The conclusions, as always, are at the end.

    Interesting little things



    Folding knobs on switches and the same ones on power supplies



    Control 2 nodes of a multi-server server from one muzzle


    Convenient replacement of coolers, all instructions are attached


    Usually the server has a whole cover, when it is opened, the server may overheat. In the case of Quanta, the server cover is divided into segments



    Latches for holding cables on power supplies

    Now let's look at the Quanta servers themselves


    There are several groups on technical features and purpose. There are also two types of additional software for server hardware:
    • Quanta System Manager (QSM) is a data center management system with an intuitive graphical interface. Essentially, Quanta's iron management system.
    • Quanta Datacenter Manager is a Quanta System Manager with additional features. It can not only control the iron, but also receive information from various sensors, make changes to the operation of external systems, such as air conditioning. Due to this, it is supposed to optimize the cost of electricity, work properly with cooling, etc.


    On sale QuantaGrid and QuantaPlex - new modern models. STRATOS - old iron, it is unlikely to be bought in Russia.


    Quanta Server Rack

    Standard Rack Servers


    Yes, there are such in the lineup. These are the options for the glands that will cover 90% of the client's needs, standard workhorses. Servers are made in 1U / 2U / 4U form factor, in rack-mount version. Built on Intel x86. A wide selection of components, such as mezzanine cards and RAID controllers, various disk configuration options (3.5 or 2.5 inch drives). For some models, it is possible to install hot-swappable NVMe PCIe SSD cards. NVMe is a special protocol that is designed to work with SSD storage media, contains a bunch of optimizations and improvements. It provides better performance and minimum response time for SSDs through the use of a new set of commands and a queuing mechanism that is optimized for working with modern multi-core processors.

    Many servers, in addition to front-panel drives, have 2–4 additional slots for SSDs. Disks can be combined in raid. They are convenient to use for the OS or for caching data. Disks are very thin (somewhere 2 times thinner than 2.5-inch drives), they are often placed on the back of the server. It looks something like this:



    GPGPU Servers


    These are the same rackmount servers. Their key difference is the presence of one or more GPU-accelerators GPU / Xeon Phi. The main function of the GPU is HPC computing. While the CPU includes several cores optimized for sequential data processing, the GPU consists of thousands of smaller and more energy-efficient cores designed to handle several tasks simultaneously. It contains more memory than the CPU, but not as fast, and a huge number of execution units. In fact, the work of the GPU is reduced to operating a huge number of independent tasks. Thus, the servers in this series are ideal for tasks such as:
    • HPC (High Performance Computing) / technical computing (various mathematical parallel / distributed computing, of course, the application code should be written taking into account the calculation of data on GPGPU);
    • 3D Rendering.


    Microservers


    This solution contains a large number of computational nodes in a relatively small number of units. At the same time, a higher density of CPU and memory per unit in the rack is achieved. It can contain up to 24 nodes in 3 units, in each node you can install 2 or 4 disks (depending on the server model) for the OS. Disks can be combined into raid. It looks something like this:


    Models on 12 and 9 nodes in 3 units.

    The server is perfect for:
    • dedicated hosting;
    • cloud computing;
    • big data processing;
    • content delivery networks;
    • as a server to host the front-end web.


    Multisite servers (cluster in a box)


    Multisite servers are 2 or 4 servers in one 2U enclosure. High density and single power supplies save space in racks and electricity for power and cooling of iron. In clustered configurations, disks in front of the server are visible to 2 nodes at a time. Servers are perfect for:
    • implementation of 2-node clusters and clusters from a larger number of nodes (search and contextual mechanisms of many highly loaded websites);
    • virtualization (VMware, Microsoft, etc.);
    • solutions for Hadoop;
    • Clients who need the maximum density of computing resources per 1U server rack;
    • energy saving.


    Storage servers


    These are servers with a maximum disk density per unit. There are many models, one essence. Here are some examples:
    • QuantaPlex T21P-4U - Holds up to 78 3.5-inch drives and 2 SSD drives for the OS. The rack occupies 4 units. The server is designed for installation in a deep rack. The rear will stick out of the standard rack.
    • QuantaGrid D51PH-1ULH is a server with 12 3.5-inch drives and 4 SSDs. In a rack it takes 1 unit. The server is slightly longer than standard servers, but all hot-plug components are replaced without being pulled out of the rack. “How to change 12 disks?” - you ask. And so, as I already showed at the very top:



    The server is in the rack, the tray is in our hands
    Nowadays, SDS solutions are becoming very popular in the data storage market, so in addition to the obvious methods of application, the servers of this line are an excellent candidate for SDS solutions.

    Storage System Elements (JBODs)


    There are few models, but they turned out to be very competitive in price. In my opinion, it is worth highlighting the JBOD QuantaVault JB4242 with 32 disks installed on the front and back of the server and the QuantaVault JB4602 with 60 disks for vertical loading. Both JBODs are made in 4U form factor.


    QuantaVault JB4602

    Network switches


    Quanta offers a rich selection of Bare Metal switches with 1G / 10G Ethernet ports. There are models with uplink 40G. The line of switches covers all customer needs: from entry-level switches with 1G ports to Top-of-Rack switches. Bare Metal is bare metal. There is no OS inside. Various options can be found on the Quanta website or downloaded from the websites of companies that develop such Firmware.
    Naturally, not all customers are interested in bare metal, so Quanta also offers models with an already pre-installed OS. Such that stuck - and it works.

    QCT Rackgo


    QCT Rackgo - a solution based on the OCP Open Cloud Server (OCS) specifications introduced by Microsoft, is innovative for providing business applications designed to integrate the server, data warehouse and network functionality with data exchange technologies and efficient management.

    It looks like a 42-unit rack with power supplies and cooling, but there is also a half option (probably it is supposed to be used for remote offices). Blades that do not have their own power supplies or coolers are inserted into the rack. There is a storage blade (JBOD), a data processing blade and network switches. In essence, this solution is one big blade of the chassis. Energy saving declared by the manufacturer is 35%.

    By the way, the manufacturer claims that the QCT Rackgo can be cooled with air from the street without preliminary cooling (in fact, take it outside and return it there). Thus, additional savings in cooling iron are possible. The solution is perfect for new projects that require a large amount of computing and storage resources, as well as for projects involving rapid growth of the hardware component.

    Equipment testing


    Several models of server hardware were delivered to CROC for testing. Here is the list:
    • QuantaPlex T21P-4U;
    • QuantaGrid D51PH-1ULH-1U Server;
    • Network switch QuantaMesh T3048-IZ1.

    In the future I will talk in more detail about each of them, but first ...

    General impressions


    Servers are made soundly, there are no agnails and dangling wires. All iron looks logical and harmonious, easy to open and disassemble. It’s nice to work with him.
    Each server has a remote management console. It can work both through a separate network port, and fumble network ports for data from the OS. This functionality allows you to reduce the number of switching inside the rack and free up network ports in the switches. In terms of functionality, the console is no different from similar consoles of other vendors. It works smartly, without brakes. It is possible to update the BIOS and IMM from the same console, and not from the boot disk.

    The OS is installed using the bootable Universal Deployment Kit. The list of supported OS surprised me:
    • Windows 2008 R2 \ 2012 \ 2012 R2;
    • Rhel 6.5 & 7;
    • Sles 11 SP3;
    • Ubuntu 14.04 & 14.04.1.


    Where, one wonders, is everything else?

    In fact, the list of supported OSs (those that can be installed on the server) is much wider, but you have to install them as they are, adding drivers after installing the OS. The manufacturer’s website has a section with drivers for Linux / Window / ESX, from where you can download everything you need. In the laboratory, we used the latest versions of Ubuntu and SLES, it worked perfectly.

    The site has very little documentation. In fact, there is nothing but a datasheet. This causes some inconvenience. To combine additional SSDs into raid, I had to write to the manufacturer. It turned out that in order to get this opportunity, you need to get into the BIOS and change a couple of settings. Then everything is obvious. How, interestingly, do people find this without documentation?
    Now in more detail for each server:

    QuantaPlex T21P-4U


    This is a 4-unit server, available in several configurations.
    • node and 78 disks;
    • 2 nodes and 70 disks. The 2nd node and tray for 8 disks are interchangeable components, each server has 35 disks available.

    Each node can be additionally supplied with 2 hot-swappable SSDs. Disks are combined in Raid1. It is convenient to place guest OS on these disks.
    The disk controller allows you to create only 2 Raid groups, which is not enough for 70 disks, so disks must be collected in Raid groups at the software level.
    The length of the server corresponds to the length of a standard rack, the sleeve is already sticking out of it. Thus, the server is designed for installation in a deep rack.
    QuantaPlex T21P-4U can be used for any tasks requiring a large number of disks. It could be:
    • archival storage;
    • CCTV;
    • file storages;
    • SDS and distributed storage (CEPH or EMC ScaleIO).


    QuantaGrid D51PH-1ULH-1U


    The server occupies 1 unit. As in the previous model, its length corresponds to the length of the rack, but all the hot-plug components of this server can be changed without removing it from the rack. Thus, he does not need a sleeve, and the server can be installed in a standard rack.

    The server has 12 3.5-inch drives and 4 SSDs. Raid controller supports various types of Raid groups. A 3.5-inch drive tray extends from the front of the server.

    The server is well suited for SDS and distributed data storage (CEPH or EMC ScaleIO), for storing and processing large data (Hadoop).

    Network Switch QuantaMesh T3048-IZ1


    For testing, in addition to server hardware, we also received a network one, namely QuantaMesh T3048-IZ1. The solution is interesting and stands out from the general system, but first things first.

    This is how the switches look in front:


    And here it is in back:


    Everything is assembled efficiently and soundly. Note the cable holders and handles on the front panel - a trifle, but nice.
    But the most interesting, of course, is inside the switch. It is built on the Intel ONP platform (Open Network Platform) - this is an open reference design for high-performance network platforms and data centers, which was laid with thoughts on SDN (Software Defined Networking) and NFV (Network Functions Virtualization).
    This means that this switch at the iron level supports OpenFlow packets and is ready to work with Open vSwitch.
    Note that the Intel ONP platform itself appeared for a reason and is an attempt to suppress Broadcom, which now feels great in the network segment.
    Since the solution is standardized, in any Intel ONP switch you will find the Intel Ethernet Switch FM6000 switching matrix and the server part based on the x86 processor. Intel Ethernet Switch FM6764 and Intel Core i3-2115C with 2 GB DDR3 and 32GB SSD work side by side in our switch. This configuration provides a throughput of 1.28 Tb / s or 960 million packet / s with a delay of <1 μs, while we get 48 ports 1 / 10GbE SFP + and 4 ports 10 / 40GbE QSFP +.
    Having talked about hardware, let's move on to the software. Intel came up thoroughly and developed a whole platform of Intel ONS (Open Network Software), which is based on Wind River Linux. This operating system is interesting in that it has a real-time kernel and advanced network interfaces with increased speed.
    Wind River itself was recently bought by Intel and deals exclusively with real-time operating systems, as an example of their work - the Curiosity rover OS, so there is no reason not to trust such an OS. SDK and API, as well as various debugging utilities have been released for the platform itself, which will allow you to modify the OS for yourself or add some additional features up to your own protocols. If this is not enough, then the open bootloader is installed on the switch itself, and thanks to the x86 architecture, any other OS is available.
    We list the available features separately:

    LevelOpportunities
    L2Port-based VLAN, 802.1Q VLAN, IGMP snooping, LACP, Storm Control, STP / RSTP / MSTP, Q-in-Q, QoS / DiffServ, L2 / L3 / L4 ACL, LLDP (802.1ab)
    L3VLAN routing, OSPF, ECMP, ARP, IGMP, PIM-SM, VRRP, OSPF, BGP
    At the data center level802.1Qaz (ETS), 802.1Qbb (PFC), DCBX, VM Tracer, EVB / 802.1Qbg, OpenFlow v1.0, VXLAN, NVGRE


    Cost


    It is clear that the price can not be considered at all and never, it is clear that the servers for different tasks are different. But I’ll comment that in specialized cases (microservers, GPU farms, disk farms) Quanta is 10-15 percent cheaper than its counterparts, in other cases at parity or slightly more expensive than the same HP / DELL / Huawei.

    Useful links:
    Tags:

    Also popular now: