Resource management is the allocation of resources from resource providers to resource consumers.

The need for resource management arises from the overcommitment of resources—that is, more demand than capacity and from the fact that demand and capacity vary over time. Resource management allows you to dynamically reallocate resources, so that you can more efficiently use available capacity.

An Example of Overcommitment

• Eight databases, each needing 400 MHz CPU at peak utilization, but significantly less during normal load.
• Two machines, each with a 1.6 GHz CPU.

The Physical Solution

1. Install 4 databases on server 1.
2. Install 4 databases on server 2.
3. Perhaps use something to restrict resource consumption (something like Oracle Instance Caging) or just hope for the best.

If you were doing this you would run without HA and worry about growth and the fact that one database could get too busy and interfere with the others.  In most production cases, you would probably need to purchase 1-2 more servers in order to run reliably.

In vSphere, there are a some ideas that are important to understand:

The Guest O/S

The guest is the combination of the applications and operating system (Windows, Linux, Mac), the same one that you may be executing today.  However, instead running on its own physical hardware, the guest sits on top of the hypervisor, which abstracts the hardware underneath.  The guest executes based on the operating platform that is presented to it, not the physical hardware itself.

A guest typically will not have any visibility to the virtualization layer, and it will not know the difference between running on physical hardware or a hypervisor.  Also, it is common that a ‘template’ can be made such that it is very easy to make many copies of a guest running with minimal effort.  This is a huge productivity savings for system and application administrators.

Thinking about databases, your guest might be an installation of Oracle running on Linux within a VMware VM.

The Host

In the physical world, a host is simply an x86 system, typically a blade of some sort in a data center.  In the virtual world, the host represents the aggregate computing and memory resources of the physical x86 server. For example, if the physical x86 server has four dual-core CPUs running at 4GHz each and 32GB of system memory, the host has 32GHz of computing power and 32GB of memory available for running virtual machines that are assigned to it.  Because there are multiple guests running on a host, it is easiest to think of a host in terms of an aggregate, as it is very easy for a hypervisor to allocate resources to fit the needs of your current workload.

As we said before, a database running on an O/S (such as Oracle on Linux) encapsulated in a VM would be considered a guest.  You could easily have a large number of databases running on the same host without interfering with one another (in fact, they would probably run better together, more on that later).  Each guest is completely isolated from one another and they will not consume each others resources, as they are managed by the hypervisor.

Also, multiple guests can be running on the same hosts and they do not need to be segregated by application type.  SQL Server, Oracle, and PostgreSQL could run on the same physical host without any issues, and a web server could easily run on the same host as a database server  as separate guests.

The Cluster

The cluster is a logical aggregate of a number of hosts, but acts and is managed as a single entity. It represents the aggregate computing and memory resources of a group of physical x86 servers sharing the same network and storage arrays. For example, if the group contains eight servers with four dual-core CPUs each running at 4GHz and 32GB of memory. The cluster then has and aggregate 256GHz of computing power and 256GB of memory available for running virtual machines.

Instead of pairing databases to specific machines, the Guest VM is deployed to a cluster.  The underlying virtual architecture will make sure that the guest is deployed to the host with the correct resource availability, and can even move guests around the cluster as needed.  VMware technologies like HA, DRS, and vMotion  (more details in Bootcamp 3) take care of the details at deployment time and during operation.

In a practical sense, the clustering technology frees databases from their physical shell, like the ghost in the machine.  In my years in development and consulting, it is rare that issues around databases crop up because of tables, indexes, or data types.  Much more often issues arise around resources and interactions between hardware, software, and drivers.  Keeping your database in a highly automated, well managed cluster allows you to get the best out of the technology without being dragged down by physical dependencies.

Resource Pools

OK, let’s think about the world from the point of view of a guest.  In the physical world, it would get its resources by querying the hardware it is running on to find out how much CPU/RAM is available for consumption.  However, in the physical world all the guest can see is what the hypervisor allows it to see.  In reality, the resources of a host can be subdivided and aggregated, almost at will.  This means that we can give a guest part or all of a given host, depending on how much it needs.

Conceptually, VMware calls these resource pools, and they are partitions of computing and memory resources from a single host or a cluster.  Resource pools can be hierarchical and nested. You can partition any resource pool into smaller resource pools to further divide and assign resources to different groups or for different purposes.

A database would receive its resources from a specific resource pool instead of from the actual physical hardware.  Keep in mind that the guest has no idea that there is anything out of the ordinary here, but resource pools can be used to manage large numbers of processes very effectively.

Overcommitment

A very common situation with databases (and other applications) is that the amount of resources that are needed at peak usage are significantly lower than the amount of resources that are needed during average usage.  Also, it is not uncommon that peak usage is only a tiny part of the actual time that a system is running.  For example, a BI database might be very busy during business hours while a data mart used for ETL might only be busy after hours. However, in the physical world you need to two under-utilized systems in order to run your architecture effectively.

However, in the virtual world we have a solution for this problem, called overcommitment.  Overcommitment of resources is the ability to provision more resources than are actually physically available.  This may seem counter-intuitive, but it stems from the fact that different applications are busy at different times.  Because the guest is abstracted from the physical hardware of the cluster, resources can be effectively ‘borrowed’ from a less busy guest and ‘loaned’ to a more busy one.   Overcommitment is done seamlessly by the hypervisor, this is done in a way that is completely automated.

In our example, we might have a physical host with 8GHz of CPU and 64GB of RAM.  However, we can assign a VM with the BI database 8GHz and 64GB of RAM AND we can assign the VM with the ETL Server 8 GHz and 64MB RAM (an overcommitment level of 100%).  Because the two systems have different peak usage, the hypervisor would be able to move resources from one VM to the other without a problem.  Even if there is some overlap, typically virtual resources can be successfully overcommitted with minimal effort.  In addition, Dynamic Resource Scheduling can transparently take the edge of any resource contention.

Why This Matters

These core concepts are very different from what members of a database team may be used to today.  However, it is critical to understand that  while the physical building blocks are the same, in a virtualized world your basic concepts of how things work can be very different.

The Takeaway

While most of these concepts are fairly straightforward, resource overcommitment is something that may be new to database people.  We will explore this in much further detail …

References

Introduction to vSphere

Memory Overcommitment

manages large collections of infrastructure (such as CPUs, storage, and networking) as a seamless and dynamic operating environment, and also manages the complexity of a datacenter.

ESX

The Takeaway

References

To quote:

 A virtual machine is a tightly isolated software container that can run its own operating systems and applications as if it were a physical computer. A virtual machine behaves exactly like a physical computer and contains it own virtual (ie, software-based) CPU, RAM hard disk and network interface card (NIC).

An operating system can’t tell the difference between a virtual machine and a physical machine, nor can applications or other computers on a network. Even the virtual machine thinks it is a “real” computer. Nevertheless, a virtual machine is composed entirely of software and contains no hardware components whatsoever. As a result, virtual machines offer a number of distinct advantages over physical hardware.

In general, VMware virtual machines possess four key characteristics that benefit the user:

• Compatibility: Virtual machines are compatible with all standard x86 computers
• Isolation: Virtual machines are isolated from each other as if physically separated
• Encapsulation: Virtual machines encapsulate a complete computing environment
• Hardware independence: Virtual machines run independently of underlying hardware

Multiple systems, one machine, no one any the wiser except for the system administrator who teaches them how to share.  From the database point of view, this is FANTASTIC, because it removes a huge number of headaches for DBAs that aren’t related to the database.

Also, because the operating system is encapsulated and machne independent, there are some important things things that you can do with VM’s that are difficult to do with physical machines.

Snapshots

A snapshot preserves the state and data of a virtual machine at a specific point in time.  The state includes the virtual machine’s power state (for example, powered-on, powered-off, suspended).  The data includes all of the files that make up the virtual machine. This includes disks, memory, and other devices, such as virtual network interface cards.

This operation is typically VERY fast compared to a database backup, typically only taking seconds.  This can be a great way to do a simple checkpoint or an incremental backup.  You just quiesce the database, take a VM snapshot and continue on your merry way.

Cloning

A full clone is an independent virtual machine, with no need to access the parent. Full clones do not require an ongoing connection to the parent virtual machine. Because a full clone does not share virtual disks with the parent virtual machine, full clones generally perform better than linked clones. However, full clones take longer to create than linked clones. Creating a full clone can take several minutes if the files involved are large.

This operation is also typically VERY fast compared to a database clone and the clone can be deployed anywhere that you want to.  While there is more work that is needed to be done inside the new VM around database configuration,  there isn’t anything that needs to be done around hardware.

Linked Clones

A linked clone is made from a snapshot of a parent virtual machines. All files available on the parent at the moment of the snapshot continue to remain available to the linked clone. Ongoing changes to the virtual disk of the parent do not affect the linked clone, and changes to the disk of the linked clone do not affect the parent.

Imagine finding a problem in production and in just a few minutes being able to hand your developer a full copy of a 1 TB database that they can query and change at will without impacting anything.  Very difficult to do in the traditional model, easy to do when virtualized.

Pretty cool stuff.

Why This Matters 

In reality virtualization wouldn’t be much more than a nice parlor trick if all you did was create individual VMs and run then on your laptop.  However, virtualizing resources really opens up a world of interesting things for databases, administration, server efficiency, and data movement.

The Takeaway

Virtualization makes running existing databases cheaper and easier.  Also, if you are clever operationally, you can manipulate the container to make the database inside run exactly the way you want it to and at the cost you want to pay.  This is the domain of virtual resource management, which we will cover in detail.  However, the first thing we need to understand is VMware’s virtualization stack.

References 

Introduction to vSphere

Snapshots

Understanding Clones

How does a Snapshot Work

1. Databases and virtualization are both great technologies.
2. DBAs typically don’t know or care much about virtualization.  It simply isn’t the way things are done.
3. Virtual Infrastructure (VI) Admins typically don’t know or care much about databases.  A payload is a payload and a VM is a black box.

However, I understand exactly how it happens.  I have been working with big data, databases, data management, and ETL for the last 10 years and I had very little interaction with virtualization or centralized IT.  Databases are the crown jewels of the IT infrastructure, and are typically managed completely seperately than the rest of IT.

The Economist has a great article about digital libraries and the effects of imprecision searching for information. Large voices have talked in great detail about the Long Tail of information on the web. That is, information doesn’t lose its value on the internet as quickly as it does elsewhere because it remains available forever. For example, an article printed in a newspaper could easily be lost in the dustbin of history simply because it can’t be easily indexed and read by the people who want to see it the most. The same information in a blog will be available through Google for the rest of eternity.

However, researchers have found that exactly the opposite is happening with scientific media. The number of links to older materials actually declined as it was moved from paper to the web. In other words, the newest, most popular material was linked to more often, creating a shorter tail than had otherwise existed. You can read the article itself for more details on the nature of the experiment itself.

I think that this study actually shows the value of inaccuracy in search. Humans don’t need less accurate information to do their job today. However, in the long term we need less accurate information in order to take us down paths we might not expect. The very act of turning pages in a book might allow some bit of information to catch the eye of the reader (either consciously or unconsciously) and take them down a path they might not otherwise have traveled. Think ask.com and iTunes cover flow vs. the bubble sort.

IMHO, this highlights the fact that the brain needs a level of accuracy in order to function, but also requires a bit of fuzziness in order to grow and thrive. This is just a part of our biology and our cognition. As Information Architects, we need to understand this in order to build the right systems for people.

According to the article:

Software installed in the refrigerated box connects the student IDs and purchase data to Horizon’s point of sale servers, which automatically track students’ prepaid account IDs, along with information on whether a student qualifies for free or discounted lunch rates based on household income. School districts get reimbursed by the government for a fraction of the discounted cost when they sell a balanced meal–under USDA rules, three of five of a bread, protein, dairy or fruit and vegetables–to low-income students.

Finally, and this is the selling point for parents who want oversight of their child’s eating habits, parents can log onto a secure Web site, called MealPayPlus, to see what their child ate for lunch, or how they snacked on any given day. They can also add money to their child’s account directly on the Web site

This is a very, very cool idea and a true win-win-win for school districts, parents, and children.  The kind of solution that really can only be done with integrated technology and cooperation between everyone involved.  The only losers here are the existing vending industry and suppliers.  They have stubbornly insisted on delivering food that maximizes their profits and tried to lock them in with exclusive contracts and payments to schools that border on bribes.

Kudos to HorizonOne for their innovation and vision!

However, the ‘new trend’ of allowing users actually provide feedback into the products that they are using has produced a professorship at MIT and been mentioned in the New York Times ::sigh::  IMHO, this is simply the realization that ignoring people is less effective than communicating with them, and less effective means less profitable.  I think the technology world should be embarassed that it is so disfunctional that this is at all new or interesting enough to study.

On a similar note, Ben Stein has some interesting thoughts about how to have a business conversation.  If you plan on doing anything other than pure solo work for the rest of your life it is mandatory reading.  Interestingly, it seems that the best conversationalists are the ones who are the most considerate and do the most listening.  Sounds like organizations might want to focus on having an actual conversation with their customers …

However, the ‘new trend’ of allowing users actually provide feedback into the products that they are using has produced a professorship at MIT and been mentioned in the New York Times ::sigh::  IMHO, this is simply the realization that ignoring people is less effective than communicating with them, and less effective means less profitable.  I think the technology world should be embarassed that it is so disfunctional that this is at all new or interesting enough to study.

On a similar note, Ben Stein has some interesting thoughts about how to have a business conversation.  If you plan on doing anything other than pure solo work for the rest of your life it is mandatory reading.  Interestingly, it seems that the best conversationalists are the ones who are the most considerate and do the most listening.  Sounds like organizations might want to focus on having an actual conversation with their customers …

[Symbio CEO Jacob] Hsu and his colleagues decided they needed a feeder program to prepare college students to work for them, so they recently established software institutes in the Harbin Institute of Technology and Shandong University, both in the coastal city of Weihai. That’s where Symbio is about to establish a new development center. Says Hsu, who grew up in San Francisco: “Other companies have university partnerships; we run the university departments.”

This isn’t something Symbio undertakes lightly. “We’re a human potential factory. We’re in the talent management business,” says Hsu. “In the next couple of years the companies that win will be the ones who manage talent the best.”

The Chinese have an abundance mentality, a positive outlook for their long-term future. Not only are they doing business successfully today, they are investing in building a generation of leaders for tomorrow. At the same time, western companies are shedding jobs and looking to outsource jobs and import labor to meet the needs of the moment. The west isn’t going to lose its edge because of quarterly profit outlooks, it is going to lose it because it lacks the vision to see the future and the audacity to put itself at the center of it!