Hurdles Towards Virtulization : What we face on its way to implementation
This post is dedicated to the current challenges that we face in the creation and use of virtualization. I am writing this even at a time when virtualization is something which is sorely missing in everyday computing and beyond for me. For instance, during the past few weeks, I had my both hard disks developing bad sectors and had a miserable experience upgrading Ubuntu on my friends' disk. We need a mechanism to seamlessly move the operating systems from one machine to another. For this, we need virtualization.
Given the current hype and hoopla over this term, many people believe that this is a recent phenomenon which is not the case. Virtualization has its roots at least 40 years before, with the concept of time sharing systems. Similarly, the IBM OS/360 also heralded this new era of 'virtual machines'(VM). In order to manage these Vms, we need a special program which is known as Hypervisor. Now the challenge with hypervisor is that it needs to provide a secure, equivalent, efficient and controlled resource environment for its guests(which are the virtual machines).
So what are the challenges ?
The first challenge that comes in my mind is the performance of the VM. If the guest is not getting enough processing, then there is no benefit out of such an arrangement. This can often happen when the guest OS performs a privileged operation such as a hardware interrupt or a system call resulting in acquisition of hardware locks. Now, as the guest has to run safely, this has to be carefully emulated by the hypervisor.
There is also a concept of time loss or clock skew. As there may be more than one OS per CPU running on a hypervisor at a given instance of time, the instructions by them would be run collectively. However, from the VM's perspective, there were only its instructions, so the instructions take longer time to execute (as other VM processes also hog the CPU cycles) than what the VM actually expects it to. Naturally, this time delay propagates to larger proportions leading to problems such as time-slicing of multiple processes at the guest OS/VM level.
The processing sharing problem also has a profound effect on the performance of the overall system. For instance, on a high priority guest, a low priority process would get precedence over a higher priority process on a low priority guest.This problem has not been solved to this date and only fringe works have been done in this regard. Perhaps this has still pertained due to the fact that a lot of OSs still are created with a single operator at a time without any restrictions.
The memory management as well as addressing are also some of the prevailing problems. Memory cache management is also a problem that rises exponentially as the number of intermediaries between the processes and hardware increase. This can be understood using a simple example. Let us consider that we have a machine which as LRU method of page catching and its guest has MRU method of caching. So as soon as the process becomes old in guest, it is discarded from the memory. However, on the host it is not deleted. So new we have wasteful data on the host memory. The other extreme of this situation is when the data is frequently required in this setup. Now, the host would start to thrash. We need to control this thrashing and ensure page faults remain constant and distributed over a period of time through 'handshaking' communication between the guest and hypervisor with regard to occurrence of faults.The final problem that plagues virtualization systems is the addressing of memory which is there due to different memory addressing schemes being followed in both the hypervisor and guest. The memory addresses are divided for processes into segments, pages and further into offsets. So every process has 2 addresses and 3 memory references, leading to extra overhead. This is avoided through a translation look-aside buffer (TLB), which generally stores the recently used segment and page addresses. But for virtualized environments, this isn't enough as in normal TLB translations, the virtual addresses are transformed into a real ones but due to 3 address levels (Actual hardware address, the 'emulated' hardware address for guest and the virtual address at the guest), there exists an overhead of extra translation. Thankfully, this is being addressed by hardware vendors through hardware support such as Intel's VT-i or AMD's NPT on one end and through utilisation of these enhancements by hypervisors such as virtualbox and vmware.
Hopefully in the near future, we can expect to see more prevalence of virtualized environments, not only in heavy duty systems, but on hardware ranging from the least expensive personal computing to mission-critical systems, given the increasing maturity of virtualization .