Skip navigation

memory management

warning: Creating default object from empty value in /var/www/vhosts/ on line 33.
Original author: 
Todd Hoff

Now that we have the C10K concurrent connection problem licked, how do we level up and support 10 million concurrent connections? Impossible you say. Nope, systems right now are delivering 10 million concurrent connections using techniques that are as radical as they may be unfamiliar.

To learn how it’s done we turn to Robert Graham, CEO of Errata Security, and his absolutely fantastic talk at Shmoocon 2013 called C10M Defending The Internet At Scale.

Robert has a brilliant way of framing the problem that I’ve never heard of before. He starts with a little bit of history, relating how Unix wasn’t originally designed to be a general server OS, it was designed to be a control system for a telephone network. It was the telephone network that actually transported the data so there was a clean separation between the control plane and the data plane. The problem is we now use Unix servers as part of the data plane, which we shouldn’t do at all. If we were designing a kernel for handling one application per server we would design it very differently than for a multi-user kernel. 

Which is why he says the key is to understand:

  • The kernel isn’t the solution. The kernel is the problem.

Which means:

  • Don’t let the kernel do all the heavy lifting. Take packet handling, memory management, and processor scheduling out of the kernel and put it into the application, where it can be done efficiently. Let Linux handle the control plane and let the the application handle the data plane.

The result will be a system that can handle 10 million concurrent connections with 200 clock cycles for packet handling and 1400 hundred clock cycles for application logic. As a main memory access costs 300 clock cycles it’s key to design in way that minimizes code and cache misses.

With a data plane oriented system you can process 10 million packets per second. With a control plane oriented system you only get 1 million packets per second.

If this seems extreme keep in mind the old saying: scalability is specialization. To do something great you can’t outsource performance to the OS. You have to do it yourself.

Now, let’s learn how Robert creates a system capable of handling 10 million concurrent connections...

Your rating: None

In computing, a hypervisor, also called virtual machine manager (VMM), is one of many hardware virtualization techniques allowing multiple operating systems, termed guests, to run concurrently on a host computer. It is so named because it is conceptually one level higher than a supervisory program. The hypervisor presents to the guest operating systems a virtual operating platform and manages the execution of the guest operating systems. Multiple instances of a variety of operating systems may share the virtualized hardware resources. Hypervisors are installed on server hardware whose only task is to run guest operating systems.

The term can be used to describe the interface provided by the specific cloud computing functionality infrastructure as a service (IaaS).[1][2]

The term "hypervisor" was first used in 1965, referring to software that accompanied an IBM RPQ for the IBM 360/65. It allowed the model IBM 360/65 to share its memory: half acting like a IBM 360; half as an emulated IBM 7080. The software, labeled "hypervisor," did the switching between the 2 modes on split time basis. The term hypervisor was coined as an evolution of the term "supervisor," the software that provided control on earlier hardware.[3][4]


Robert P. Goldberg classifies two types of hypervisor:[5]

  • Type 1 (or native, bare metal) hypervisors run directly on the host's hardware to control the hardware and to manage guest operating systems. A guest operating system thus runs on another level above the hypervisor.
This model represents the classic implementation of virtual machine architectures; the original hypervisor was CP/CMS, developed at IBM in the 1960s, ancestor of IBM's z/VM. A modern equivalent of this is the Citrix XenServer, VMware ESX/ESXi, and Microsoft Hyper-V hypervisor.
  • Type 2 (or hosted) hypervisors run within a conventional operating system environment. With the hypervisor layer as a distinct second software level, guest operating systems run at the third level above the hardware. KVM and VirtualBox are examples of Type 2 hypervisors.

In other words, Type 1 hypervisor runs directly on the hardware; a Type 2 hypervisor runs on another operating system, such as FreeBSD[6] or Linux[7].

Note: Microsoft Hyper-V (released in June 2008)[8] exemplifies a type 1 product that can be mistaken for a type 2. Both the free stand-alone version and the version that is part of the commercial Windows Server 2008 product use a virtualized Windows Server 2008 parent partition to manage the Type 1 Hyper-V hypervisor. In both cases the Hyper-V hypervisor loads prior to the management operating system, and any virtual environments created run directly on the hypervisor, not via the management operating system.


Mainframe origins

The first hypervisor providing full virtualization, IBM's one-off research CP-40 system, began production use in January 1967, and became the first version of IBM's CP/CMS operating system. CP-40 ran on a S/360-40 that was modified at the IBM Cambridge Scientific Center to support Dynamic Address Translation, a key feature that allowed virtualization. Prior to this time, computer hardware had only been virtualized enough to allow multiple user applications to run concurrently (see CTSS and IBM M44/44X). With CP-40, the hardware's supervisor state was virtualized as well, allowing multiple operating systems to run concurrently in separate virtual machine contexts.

Programmers soon re-implemented CP-40 (as CP-67) for the IBM System/360-67, the first production computer-system capable of full virtualization. IBM first shipped this machine in 1966; it included page-translation-table hardware for virtual memory, and other techniques that allowed a full virtualization of all kernel tasks, including I/O and interrupt handling. (Note that its "official" operating system, the ill-fated TSS/360, did not employ full virtualization.) Both CP-40 and CP-67 began production use in 1967. CP/CMS was available to IBM customers from 1968 to 1972, in source code form without support.

CP/CMS formed part of IBM's attempt to build robust time-sharing systems for its mainframe computers. By running multiple operating systems concurrently, the hypervisor increased system robustness and stability: Even if one operating system crashed, the others would continue working without interruption. Indeed, this even allowed beta or experimental versions of operating systems – or even of new hardware[9] – to be deployed and debugged, without jeopardizing the stable main production system, and without requiring costly additional development systems.

IBM announced its System/370 series in 1970 without any virtualization features, but added them in the August 1972 Advanced Function announcement. Virtualization has been featured in all successor systems. (All modern-day (as of 2009[update]) IBM mainframes, such as the zSeries line, retain backwards-compatibility with the 1960s-era IBM S/360 line.) The 1972 announcement also included VM/370, a reimplementation of CP/CMS for the S/370. Unlike CP/CMS, IBM provided support for this version (though it was still distributed in source code form for several releases). VM stands for Virtual Machine, emphasizing that all, and not just some, of the hardware interfaces are virtualized. Both VM and CP/CMS enjoyed early acceptance and rapid development by universities, corporate users, and time-sharing vendors, as well as within IBM. Users played an active role in ongoing development, anticipating trends seen in modern open source projects. However, in a series of disputed and bitter battles, time-sharing lost out to batch processing through IBM political infighting, and VM remained IBM's "other" mainframe operating system for decades, losing to MVS. It enjoyed a resurgence of popularity and support from 2000 as the z/VM product, for example as the platform for Linux for zSeries.

As mentioned above, the VM control program includes a hypervisor-call handler which intercepts DIAG ("Diagnose") instructions used within a virtual machine. This provides fast-path non-virtualized execution of file-system access and other operations. (DIAG is a model-dependent privileged instruction, not used in normal programming, and thus is not virtualized. It is therefore available for use as a signal to the "host" operating system.) When first implemented in CP/CMS release 3.1, this use of DIAG provided an operating system interface that was analogous to the System/360 SVC ("supervisor call") instruction, but that did not require altering or extending the system's virtualization of SVC.

In 1985 IBM introduced the PR/SM hypervisor to manage logical partitions (LPAR).

UNIX and Linux servers

Several factors led to a resurgence around 2005[10] in the use of virtualization technology among UNIX and Linux server vendors:

  • expanding hardware capabilities, allowing each single machine to do more simultaneous work
  • efforts to control costs and to simplify management through consolidation of servers
  • the need to control large multiprocessor and cluster installations, for example in server farms and render farms
  • the improved security, reliability, and device independence possible from hypervisor architectures
  • the ability to run complex, OS-dependent applications in different hardware or OS environments

Major UNIX vendors, including Sun Microsystems, HP, IBM, and SGI, have been selling virtualized hardware since before 2000. These have generally been large systems with hefty, server-class price-tags (in the multi-million dollar range at the high end), although virtualization is also available on some mid-range systems, such as IBM's System-P servers, Sun's CoolThreads T1000, T2000 and T5x00 servers and HP Superdome series.

Multiple host operating systems have been modified[by whom?] to run as guest OSes on Sun's Logical Domains Hypervisor. As of late 2006[update], Solaris, Linux (Ubuntu and Gentoo), and FreeBSD have been ported to run on top of Hypervisor (and can all run simultaneously on the same processor, as fully virtualized independent guest OSes). Wind River "Carrier Grade Linux" also runs on Sun's Hypervisor.[11] Full virtualization on SPARC processors proved straightforward: since its inception in the mid-1980s Sun deliberately kept the SPARC architecture clean of artifacts that would have impeded virtualization. (Compare with virtualization on x86 processors below.)[12]

HP calls its technology to host multiple OS technology on its Itanium powered systems (Integrity) "Integrity Virtual Machines" (Integrity VM). Itanium can run HP-UX, Linux, Windows and OpenVMS. Except for OpenVMS, to be supported in a later release, these environments are also supported as virtual servers on HP's Integrity VM platform. The HP-UX operating system hosts the Integrity VM hypervisor layer which allows for many important features of HP-UX to be taken advantage of and provides major differentiation between this platform and other commodity platforms - such as processor hotswap, memory hotswap, and dynamic kernel updates without system reboot. While it heavily leverages HP-UX, the Integrity VM hypervisor is really a hybrid that runs on bare-metal while guests are executing. Running normal HP-UX applications on an Integrity VM host is heavily discouraged[by whom?], because Integrity VM implements its own memory management, scheduling and I/O policies that are tuned for virtual machines and are not as effective for normal applications. HP also provides more rigid partitioning of their Integrity and HP9000 systems by way of VPAR and nPar technology, the former offering shared resource partitioning and the later offering complete I/O and processing isolation. The flexibility of virtual server environment (VSE) has given way to its use more frequently in newer deployments.[citation needed]

IBM provides virtualization partition technology known as logical partitioning (LPAR) on System/390, zSeries, pSeries and iSeries systems. For IBM's Power Systems, the Power Hypervisor (PowerVM) functions as a native (bare-metal) hypervisor and provides EAL4+ strong isolation between LPARs. Processor capacity is provided to LPARs in either a dedicated fashion or on an entitlement basis where unused capacity is harvested and can be re-allocated to busy workloads. Groups of LPARs can have their processor capacity managed as if they were in a "pool" - IBM refers to this capability as Multiple Shared-Processor Pools (MSPPs) and implements it in servers with the POWER6 processor. LPAR and MSPP capacity allocations can be dynamically changed. Memory is allocated to each LPAR (at LPAR initiation or dynamically) and is address-controlled by the POWER Hypervisor. For real-mode addressing by operating systems (AIX, Linux, IBM i), the POWER processors (POWER4 onwards) have architected virtualization capabilities where a hardware address-offset is evaluated with the OS address-offset to arrive at the physical memory address. Input/Output (I/O) adapters can be exclusively "owned" by LPARs or shared by LPARs through an appliance partition known as the Virtual I/O Server (VIOS). The Power Hypervisor provides for high levels of reliability, availability and serviceability (RAS) by facilitating hot add/replace of many parts (model dependent: processors, memory, I/O adapters, blowers, power units, disks, system controllers, etc.)

Similar trends have occurred with x86/x86_64 server platforms, where open-source projects such as Xen have led virtualization efforts. These include hypervisors built on Linux and Solaris kernels as well as custom kernels. Since these technologies span from large systems down to desktops, they are described in the next section.

PCs and desktop systems

Interest in the high-profit server-hardware market sector has led to the development of hypervisors for machines using the Intel x86 instruction set, including for traditional desktop PCs. One of the early PC hypervisors, the commercial-software VMware, debuted in 1998.

The x86 architecture used in most PC systems poses particular difficulties to virtualization. Full virtualization (presenting the illusion of a complete set of standard hardware) on x86 has significant costs in hypervisor complexity and run-time performance. Starting in 2005, CPU vendors have added hardware virtualization assistance to their products, for example: Intel's Intel VT-x (codenamed Vanderpool) and AMD's AMD-V (codenamed Pacifica). These extensions address the parts of x86 that are difficult or inefficient to virtualize, providing additional support to the hypervisor. This enables simpler virtualization code and a higher performance for full virtualization.

An alternative approach requires modifying the guest operating-system to make system calls to the hypervisor, rather than executing machine I/O instructions which the hypervisor then simulates. This is called paravirtualization in Xen, a "hypercall" in Parallels Workstation, and a "DIAGNOSE code" in IBM's VM. VMware supplements the slowest rough corners of virtualization with device drivers for the guest. All are really the same thing, a system call to the hypervisor below. Some microkernels such as Mach and L4 are flexible enough such that "paravirtualization" of guest operating systems is possible.

In June 2008 Microsoft delivered a new Type-1 hypervisor called Hyper-V (codenamed "Viridian" and previously referred to as "Windows Server virtualization"); the design features OS integration at the lowest level.[13] Versions of the Windows operating system beginning with Windows Vista include extensions to boost performance when running on top of the Hyper-V hypervisor.

Embedded systems

As of 2009[update] virtual machines have started to appear in embedded systems, such as mobile phones. This provides a high-level operating-system interface for application programming, such as Linux or Microsoft Windows, while at the same time maintaining traditional real-time operating system (RTOS) APIs. The low-level RTOS environments need to be retained for legacy support, and because the real-time capabilities of high-level OSes are insufficient for many embedded applications.

Embedded hypervisors must therefore have real-time capability, a design criterion not present for hypervisors used in other domains. The resource-constrained nature of many embedded systems, especially battery-powered mobile systems, imposes a further requirement for small memory-size and low overhead. Finally, in contrast to the ubiquity of the x86 architecture in the PC world, the embedded world uses a wider variety of architectures. Support for virtualization requires memory protection (in the form of a memory management unit or at least a memory protection unit) and a distinction between user mode and privileged mode, which rules out most microcontrollers. This still leaves x86, MIPS, ARM and PowerPC as widely deployed architectures on medium- to high-end embedded systems.

As manufacturers of embedded systems usually have the source code to their operating systems, they have less need for full virtualization in this space. Instead, the performance advantages of paravirtualization make this usually the virtualization technology of choice. Nevertheless, ARM has recently added full virtualization support as an IP option and has included it in their latest high end processor codenamed Eagle.

Other differences between virtualization in server/desktop and embedded environments include requirements for efficient sharing of resources across virtual machines, high-bandwidth, low-latency inter-VM communication, a global view of scheduling and power management, and fine-grained control of information flows.[14]

Security implications

The use of hypervisor technology by malware and rootkits installing themselves as a hypervisor below the operating system can make them more difficult to detect because the malware could intercept any operations of the operating system (such as someone entering a password) without the antivirus software necessarily detecting it (since the malware runs below the entire operating system). Implementation of the concept has allegedly occurred in the SubVirt laboratory rootkit (developed jointly by Microsoft and University of Michigan researchers[15]) as well as in the Blue Pill malware package. However, such assertions have been disputed by others who claim that it would indeed be possible to detect the presence of a hypervisor-based rootkit.[16]

In 2009, researchers from Microsoft and North Carolina State University demonstrated a hypervisor-layer anti-rootkit called Hooksafe that can provide generic protection against kernel-mode rootkits.[17]

See also

Your rating: None

StormDriver writes "According to writer Marc Prensky, most of us come from a generation of digital immigrants. It basically means the modern web developed during our lifetime, it is a place we migrated to, discovering its potential. But people aged 20 and younger are not like that at all. They are digital natives, they've spent their whole lives here. 'Hey, let's do a digital version of our college facebook' is a digital immigrant's idea, just like 'Hey, let's make something like a classifieds section of a newspaper, only this one will be online.' Or 'Hey, let's make an online auction housel.' 'Hey, let's make a place for online video rentals.' The thing is, recreating items, ideas and interactions from the physical realm on the Web already ran its course." To me, this sounds like the gripe that "Everything that can be invented, has been invented." There are a lot of real-life services and experiences that have yet to be replicated, matched, or improved upon in the online realm; I wouldn't want people to stop taking inspiration from "old fashioned" goods as starting points for digital products.

Share on Google+

Read more of this story at Slashdot.

Your rating: None

We're looking for an outstanding Network Software Engineer to join our team! As a Network Software Engineer you'll connect the games we make to exciting online play-spaces. You'll design high-level protocols and implement them in support of gameplay features. You will also optimize low-level networking systems for games that must work given bandwidth, latency, and packet loss specifications. Join our team as the official Destroyer of Lag!

Your rating: None