Category Archives: cloud

Choosing A Cloud Software Partner

Posted on by
Reply

This is the third in a series of four articles discussing infrastructure as a service (IaaS) clouds. The series started at basic level setting and we will now begin diving progressively deeper. The topics for the series are:

1. Cloud 101
    – What is cloud
    – What value should cloud provide
    – Public, private, and hybrid cloud
    – Starting on a cloud project
2. Application taxonomy, what belongs in the cloud, and why
3. What you should look for in cloud infrastructure software
4. Evaluating different approaches to cloud infrastructure software

The concepts section describes architectural design points you should ask vendors about to make sure that the they are thinking like a true cloud provider and not simply cloud-washing older technology to try to be relevant in a new world. Keep in mind that these concepts are about infrastructure management in general, not just compute. You should think about the storage, network, and power aspects of your cloud in the same way.

The specific functionality section lists cloud features to check for.  If too many of these are missing, the cloud value proposition will not be delivered.

Core Concepts and Philosophy

Scale

In a cloud, scale is the key to long-term success. The number of nodes and instances, simultaneous connections to the management system, the networking and security features, etc. all need to scale. For each and every exciting and valuable feature a cloud vendor touts, you need to ask, “Can I have tens of thousands of those? What is the experience at that scale? When, if ever, does the scale impact the end user and how they do their work?”

While one can deploy a small-scale cloud, if a cloud is successful, it will become a single pool of capacity for an entire organization or even multiple organizations. In fact, the larger clouds scale, the more cost savings and value they generate since you start to see the benefits of  “the law of large numbers”. If you do not build for scale from the very beginning, you will hit a wall and need to create separately managed clouds. This will force end users to decide what workloads go on what clouds, thus the frictionless self-service model is broken. Furthermore, capex benefits will be lost as you are forced to overprovision each cloud fragment rather than benefiting from a single pool hosting many applications with offsetting resource consumption curves.

For example, in a non-cloud deployment, the datacenter management system is used by datacenter admins only. If it can only deal with tens of simultaneous connections and is limited to one or two nodes, there is no problem since the administrative team is relatively small. However, in a cloud, since a large number of end users drive the management system directly via self-service workflows, the management system requires a whole new level of scale.

Automation

Automation is the key for allowing end users to do their own work and also for lowering datacenter operation costs. Make sure there is a proper degree of automation for both application lifecycle operations and infrastructure operations.

The core principle for end user operations is that no end user task should ever trigger work on the datacenter administrator side, not even a single mouse click approval. There is still a high degree of control and protection required, but these controls must be implemented as up front policies where the right groups of people delegate the right privileges to the right consumers. Furthermore, there need to be audit trails so that one can show that the policies constrained people to the proper activities. However, none of this changes the fact that manual approval processes by central admins on regular daily end user operations cannot work in a cloud model.

The core principle for datacenter operations is that the cloud should be self-discovering, self-organizing, self-monitoring, and self-healing. Anyone that sells you a complete zero-touch datacenter today is certainly exaggerating, but you should check the features they have to make sure this philosophy is followed where technically feasible. Where manual intervention is needed, make sure that this intervention is required only for infrequent up front tasks and never for frequent operations that happen on a frequent basis.

As an example, initial cloud configuration and network setup may be items that require significant up front planning and hours to days of setup, but regularly growing the cloud deployment by adding nodes must require no more time and effort than it takes to rack the systems and plug them in.

Identity, Permissions and Delegation

Clouds need to understand who each user is, what groups they belong to and what customer or tenant their work is billed to. Each operation on each object needs to check the identity of the actor against the permissions system to make sure that the operation is allowed. Delegation then needs to be possible – from cloud admin to customer admin to end users and groups, and possibly between separate end users and groups. Without a strong concept of identity, permissions and delegation, your cloud will only scale to a single tenant and will never fully interoperate well with other clouds, thereby limiting the long-term benefit you derive from the system. Like scale and automation, this is a core design choice.

If cloud vendors do not have proper permissions systems for their objects or are lacking a way to delegate permissions through multiple levels, they are not thinking like a cloud vendor. The result will be trouble down the road as end users wind up having to place tickets to acquire permissions driving a heavyweight approval process where the owner of the resource and the end user’s management team need to be consulted.

Openness and Choice

Openness and choice mean that you have:

  • Independence at each layer: Your different cloud components are not locked in from end to end. A choice at one layer does not dictate a choice at another unrelated area.

    • Your choice of end-user self-service workflow management should never dictate your hypervisor or other core infrastructure component.
    • Equally importantly, your private cloud software should never dictate the choice of public clouds to which can federate. Your end-user provisioning interface should work on your private cloud infrastructure, any public cloud using the same cloud software, and even any public cloud that uses competing or homegrown cloud software. Having to present different interfaces to your end users for clouds using different cloud infrastructure components is not open.
  • Complete and open APIs: Your cloud vendor should have extensive APIs. At the very least the APIs should cover everything provided in the UI. This will allow customized workflows at both the infrastructure level and the end-user level.
  • Extensible components: Your cloud vendor should use open and extensible components where possible.  Open source, where anyone can insert code at any point, is the extreme example of this principle. In non-open source systems, there are ways to introduce more controlled, but still extremely flexible extensibility models.  For example, major components can be general purpose enough that customers can add in other ecosystem products readily, as with the Linux domain 0 model for hypervisors. Alternately, APIs can be made to be robust and complete enough so that most conceivable useful integrations are possible such as the case with Windows APIs. This makes a big difference as you try to augment your cloud with best of breed 3rd party cloud management products.
  • Standards: Your cloud vendor should take advantage of open standards where possible where those standards do not unduly constrain innovation.

Without openness and choice you risk vendor lock-in and the high cost that comes from not being able to have a meaningful option to replace an infrastructure component. Technology lock-in slows down the rate at which you get new features you request from your vendor. A limited ecosystem and an inability to augment your cloud with the latest and greatest offerings from companies both new and established or from the open source community further limits your ability to improve your cloud over time. Lastly, limited choice in public clouds to which you can federate may force you into a cloud with the wrong feature set or that is too expensive.

Key Functionality

When evaluating the base functionality described here, make sure to bring in the philosophies above for each. Make sure that every feature below is implemented with scale, automation, permissions and delegation, and openness in mind.

In the spirit of openness, it is key to recognize that it is not required, nor even desirable, for all the features here to come from the cloud software vendor. Cloud software vendors should be able to present you with an ecosystem that helps fulfill the requirements below.

Self-Service Developer and Deployment Workflows

  • This is the core of the concept of cloud. End users need a way to do the following on their own:
  • Manage their images, update, and version them.
  • Publish images to a selected community for use in deployment workflows.
  • Deploy images and configure the following runtime parameters:
    • The number of instances
    • The images used
    • The placement policy
    • The network connectivity
    • The application configuration
    • The resource allocation
    • The storage to mount
  • Scale applications up and down and retire them when their usefulness has ended.

Reliability and Scale of the Management System

With cloud, when we think about the management system, we’re not just talking about basic monitoring. We’re also talking about the whole datacenter control system – how workloads are deployed, managed, and retired. In a non-cloud datacenter, the management system is for datacenter admins only. If it can only deal with tens of simultaneous connections and is limited to one or two nodes creating a single point of failure compromising reliability, there is no problem since the administrative team is relatively small. However, in a true cloud, with end users driving the management system directly in self-service workflows, a whole new level of scale is required.

The management system of a cloud needs to be able to scale to handle thousands of simultaneous connections.  Furthermore, it can never be down. It needs to be self-monitoring and self-healing. When and if a management node is lost, the remaining management nodes need to continue operation, and the lost management node needs to be replaced from the remaining equipment in the cloud so that the degraded state is resolved. All of this needs to happen automatically without impact to the end user or intervention on the part of the administrator.

Multi-Tenancy and Networking

For clouds to be of use to anything more than the smallest organizations, robust and secure multi-tenancy separation is required. This involves a great deal of network functionality.

Some customers will require traditional separation at layer 2 like VLANs. Those networks need to be managed flexibly and securely.

  • The cloud needs to allow for the layer 2 network to be exposed to large sets of nodes without difficult configuration.
  • The cloud needs to make sure that there is a security system governing access to each layer 2 network so that only the right workloads from the right customers are placed on a given network.
  • The Layer-2 network should provide the equivalent of a broadcast domain and support all traffic types (unicast, muliticast and broadcast). It should also support IPv6 and any other layer 3 protocol, not just IPv4.

Other customers, who do not want to be limited to the scale and manageability of layer 2 networks and that do not need a broadcast domain may choose to have a more modern and flexible large flat cloud network with an integrated distributed firewall providing the isolation between customers’ workloads. This distributed firewall service needs:

  • To have a central configuration repository that ultimately informs the separation between workloads created by different customers as well as between workloads within individual customers.
  • To be configurable by the infrastructure administrator, the individual customer administrators, and even the end users who need to control access to their own work product within their organizations.
  • To have its configuration be independent of workloads and IPs – adding and removing workloads must not cause reconfiguration of the distributed firewall service.
  • To execute in a distributed manner that avoids network bottlenecks.
  • To be independent of server, building, network vendor, network topology, and even geography.

Storage Management

Like compute, storage needs to be aggregated into large pools for access by end users so that they are hidden from the details of the different storage devices and what objects are placed on what storage device.

Unlike compute, there are widely varying capabilities and prices for different storage devices, so some aggregation system is needed to create pools with different service levels where customers can decide the capabilities they need and are wiling to pay for to store their storage objects. This customer decision then drives automated pool selection and ultimately, device selection.

Like with all other cloud resources, the end-user created storage objects need to be created through self-service workflows without administrator interaction, but must also be governed by a robust permission and delegation system governing which storage can be used by which users.

The storage objects created by end-users in those pools need to be managed independently of the instances that mount and access them. This way, creating, updating or deleting workloads does not affect the core information the customer needs to preserve over time. Workloads that create data can be killed, redeployed from an updated template and reattached to storage without impact to the storage object itself. Also, storage objects should be able to be cloned or snapshotted for use by future instances or for rollback processes without any interaction with the running instance accessing it.

Billing and Chargeback

Core to the economic model of cloud is the ability to have end customers either pay for their usage or at the very least to understand their impact on datacenter costs. To that end, there needs to be complete metering APIs and a chargeback or showback system for the cloud.

Hands-Off Infrastructure Management

Management of the physical infrastructure should be as low touch as possible. This includes many aspects:

  • Installation of the nodes: Node installation becomes a frequent operation in a big, fast-growing, and/or mature cloud where parts need to be replaced regularly. Manually installing or configuring servers will be too expensive and error-prone in this world. The only proper experience is for the servers to be racked and connected, then powered on – and nothing else. The cloud needs to auto-discover the server, install it, and make it ready to accept workloads.
  • Intelligent workload placement: Workloads should be automatically, and without administrator involvement, placed such that they are:
    • Loosely packed enough that bottlenecks and performance problems are not generated as dealing with those problems reactively will be problematic at scale.
    • Tightly packed enough that hardware, power, and cooling are not wasted.
    • Strategically placed so that related workloads cohabitate for enhanced inter-workload communication and that redundant workloads are separated to eliminate single points of failure for the service.
    • Placed based on constraints such as the requirement to be on a node with GPUs or a node that is certified PCI compliant.
  • Capacity tracking: There needs to be cloud-wide tracking of resources so that the datacenter operators are aware of cloud capacity and when they need to acquire more hardware.
  • Isolating and retiring equipment: All systems should have a lifetime and a health status associated with them (due to length of maintenance contract, expected lifetime of component parts, and/or length of lease). When that lifetime is exceeded or when the part is failing or has failed, it is automatically isolated from the cloud and flagged for replacement. The cloud should be aware of datacenter layout so that administrators never have a problem locating the equipment at replacement time.
  • Managing planned and unplanned downtimes within the datacenter: If you generally deploy cloud-ready applications (see last article in this series), most datacenter events should be transparent to the end users of the services. Scale-out applications can be scaled up to repopulate lost instances and chunks of huge compute jobs can be automatically respun. However, downtimes associated with persistent data need to be managed as well as compute or network downtimes that affect any of your more monolithic applications. The datacenter should recover whatever it can on its own, and for what it can’t, end users need to be alerted to upcoming planned downtime or recent unplanned downtime and made capable of adjusting their workload deployments accordingly.

Federation Across Clouds

To provide a single interface for all end-users, a cloud must hide distinctions between different datacenters, geographies and providers. There should be one end-user experience for deploying to anywhere in the cloud – public, private, or hybrid. While end users, due to compliance reasons, may need to dictate placement policy in terms of location or provider, it should be a policy component of their work within a single experience – it is not acceptable for there to be a private cloud experience and completely separate public cloud experience.

It is critical that the choice of public clouds to federate to is not forced by the cloud software provider, that would be too limiting. By not allowing the customer to pick the best possible provider at the right cost, cloud deployment will become unnecessarily expensive.

Federation features need to be as follows:

  • From a single user interface, resources can be deployed and managed across multiple sites and providers.
  • Each site and provider can be allowed or denied on a per customer or per user/group basis.
  • When accessing public clouds that are tied to shared credential and billing information, like private keys and credit card numbers, the end user must have that information hidden from them. They don’t need to know it and they must not be able to take it with them when they change jobs or roles.
  • Identity is preserved across sites and providers so that:
    • Users are permitted access to resources at each site according to their specific permissions.
    • Bills from public cloud providers can be itemized by department, user, and project.
  • There needs to be a single audit trail showing who did what activity in what clouds.

When a cloud management system follows these rules, multiple sites within an organization can be managed as one. Furthermore, hybrid cloud becomes a reality with public cloud becoming a viable part of the IT toolkit, not a bootleg process hidden from the visibility of those most trained and responsible for keeping services safe and secure.

Conclusion

Enterprise IT departments and service providers have no shortage of choices today for cloud infrastructure software. But, for an organization doing a significant deployment, the list of requirements above can help separate the serious contenders from less mature or well thought through products.

When writing RFP’s don’t just fall back on the same enterprise management or virtualization platform requirements or you will wind up with the same old infrastructure. Start with your traditional requirements, but make sure to add in critical new requirements around what is really needed to take the next step and have a real cloud today!

What Is Best Suited For The Cloud

Posted on by
Reply

This is the second in a series of four articles discussing infrastructure as a service (IaaS) clouds. The series started at basic level setting and we will now begin diving progressively deeper. The topics for the series are:

1. Cloud 101
- What is cloud
- What value should cloud provide
- Public, private, and hybrid cloud
- Starting on a cloud project
2. Application taxonomy, what belongs in the cloud, and why
3. What you should look for in cloud infrastructure software
4. Evaluating different approaches to cloud infrastructure software

Cloud is obviously a serious transformation for your datacenter, but that transformation does not need to be far off or futuristic. If you pick the right applications owned by the right users with the right needs and if you partner with the right cloud software provider, cloud and its many benefits are achievable today.

When first deploying a cloud, it is critical to choose the right applications to move to the cloud initially. Given that cloud is about allowing business units to manage their own computing needs, it follows that the ideal place for cloud is where the business units:

  • Expect self-service.
  • Are tolerant of incorporating provisioning logic into their day-to-day work.
  • Have variable compute needs for the tasks that generate the need for frequent provisioning and de-provisioning activities.
  • Have multiple tasks that cause context switching in the customers’ work, with different tasks acquiring and yielding compute resources over time.
  • Have workloads that, even at maximum interaction, do not cause complex resource contention issues on shared resources so that as much or as little can be deployed as necessary with little if any forethought or calculation.

If you break down the types of workloads in a datacenter, you get three major types:

  • Traditional, monolithic, and stateful client/server apps – things like Exchange Servers and traditional databases
  • Scale-out load balanced apps with disposable stateless instances
  • Batch type computing jobs that can be decomposed into small chunks of compute and storage and distributed across a pool – things like Hadoop, Monte Carlo simulations, business analytics, and media processing and conversion

For each class of application, there are dev/test deployments and production deployments. This gives us a simple six-way classification of what runs in a datacenter that we can use to select our cloud candidates.

The following chart shows which workloads are good for early cloud adoption and what should be dealt with later on in the process. Beneath the chart is some explanation and justification for this assessment.

(Gray represents near-term opportunity and orange represents something that should be sent to the cloud later on)

Scale-out Load Balanced Apps with Disposable Instances

This type of application does not rely on any one instance of software being able to grow to use tremendous amounts of compute resource. Rather it assumes that each instance can only do so much and that additional power will be supplied by adding more instances. For this sort of application to work the data and other application state must be driven out of the application itself to another location.This allows instances to be created and destroyed with no data loss or outage to the end user. The worst problem caused by failure of an instance is the need for the end user to retry their operation. When the operation is retried, it finds a live instance and completes without difficulty.

Generally, the application requires greater or lesser instances over time as load grows and shrinks. This creates a need for dynamism in the deployment with low turn-around time on provisioning operations. As load grows, the application must respond in a short period of time. Either the application administrator or some auto-scaling management system needs to be able to make the required changes without a ticket to the infrastructure team. When load is high, new instances must be spawned to handle the increased demands. When load drops, instances should be deleted to reclaim compute resources for more useful activity. Because the instances are stateless, without sufficient load that needs to be serviced, no instance has an inherent reason to remain persistently deployed.

Scale-out applications are also a great fit with the datacenter operations models required to manage a massive cloud.  Cloud scale datacenters need to assume that any piece of equipment can break at any time with the application being resilient and able to hide the failure from the end user. Scale-out applications accomplish this by spreading many instances across nodes, datacenters, or even geographies. This makes it much simpler for the infrastructure to be managed – failures become a capacity issue which can be managed in aggregate on a periodic basis, not end-user outage issues that need to be addressed immediately and individually.

More generally, this type of application is the way of the future in general. Due to current systems architectures, we are seeing more and more cheaper cores distributed across many commodity nodes rather than the massive scaling up of individual servers. The only way to really get applications to scale is to build for many smaller instances teaming up together. Some programming infrastructures take this to a logical conclusion. Node.js, a language increasingly used for scale-out web applications, refuses to give a program access to more than one core regardless of how many cores are in a system. Should the developer need more power, they need to increase the number of node.js application instances.

Scale-out applications and cloud are such a good match because they have the same goals – elastically, scale up and down in response to load using only the resources actually needed, distributing applications across hosts and sites to better tolerate systems outages without end user impact and better conformation to modern system architecture.

Batch Applications

Batch applications are decomposable to smaller compute and storage packages. They are good in both test/dev and production for similar reasons as the scale-out applications. As jobs are launched, the number of instances depends on how fine grained you can chunk the job. Given different sizes of data sets for separate runs, the number of instances needed for each run will vary. Therefore, statically provisioning instances doesn’t make sense. Furthermore, there are likely to be completely different applications in such an environment that will need to run at different times. Clouds are perfect for repurposing an infrastructure, ramping up one application while ramping down another without having to do a massive retooling of the physical infrastructure. Like with the last case, the infrastructure administrators should not decide how many and when to deploy instances. Activity needs to be driven by the business that is actually running the jobs and deriving benefit from the output.

Traditional IT Applications

Traditional IT applications, vertically scaled, stateful and monolithic are not good for clouds in their production deployments. These applications tend to be custom built for a particular load in mind, deployed once, expected to have each instance stay running persistently, and are only touched for upgrade. These types of applications are not a natural fit for the cloud paradigm. They do not accommodate dynamic scaling by simply adding more instances. As a result there is a reduced need for end-user self-service.  Also, since they are monolithic, any performance or availability problem must be addressed in the one or two instances that make up the application with a deep understanding of the impact of the underlying infrastructure. As a result, the burden of management remains on the infrastructure administrator and cannot shift the application administrator. This breaks the operational model of cloud where the datacenter administrator needs to be removed from the details of the running of applications.

However, traditional IT applications are very well suited to cloud when the service is development, integration, and testing of the applications. For this function, application owners need to deploy new instances of server software, update them, make new templates, try out their work and iterate. Furthermore, mass numbers of clients will need to be deployed for load and scale testing. The development and testing process generates the dynamism in workload resource needs as well as the requirement for self-service on the part of the IT developer that make a cloud an ideal solution.

Pick the Right Applications and Move to Cloud Today!

Too many times cloud is pitched as an evolutionary technology. In many cases, this is because the vendor making this pitch is already managing a legacy application stack for the customer and sees no reason for a radical shift.

Since these legacy applications do not accommodate elasticity and do not tolerate the more unpredictable availability of any single server that the cloud datacenter operations model implies, true clouds are limited in the benefits they can provide and cause a loss of SLA that is unacceptable to the end users of the legacy applications.

Of course, legacy applications will not go away any time soon and we acknowledge that it takes tremendous time and effort to move to a new programming paradigm. But, the technology is here today and the benefits have been made obvious – scale, resiliency, efficiency. The success stories of companies like Netflix and Zynga are well known. All that is needed is the will to move in that direction.

For enterprises and service providers that leverage modern applications development process, cloud is not an evolution at all – cloud is the best and most obvious way forward for development, testing and mass deployment of their applications.

Pick your target applications and get started today!

Cloud 101

Posted on by
1

This is the first in a series of four articles discussing infrastructure as a service (IaaS) clouds. The articles will start with basic level setting and will dive progressively deeper as the series progresses. The topics for the series will be:

1. Cloud 101

  • What is cloud
  • What value should cloud provide
  • Public, private, and hybrid cloud
  • Starting on a cloud project

2. Application taxonomy, what belongs in the cloud, and why
3. What you should look for in cloud infrastructure software
4. Evaluating different approaches to cloud infrastructure software

What Is Cloud?

Cloud is fundamentally about creating a dynamic computing infrastructure that enables end users to service and manage themselves in a process that is frictionless and instantaneous, but also secure and controlled. Resources are allocated in a very fine-grained fashion and can be relinquished at any time. Usage and chargeback is measured per customer, per resource utilization, and per unit of time, not by physical piece of equipment.

Cloud is not an incremental step on top of virtualization. While virtualization is a key enabler of cloud, the motivation, focus, target applications, and evaluation criteria are very different.

For clouds to be useful and cost effective, they must also have the properties of being very scalable and smooth to manage at the infrastructure level. The idea is to remove as much as possible from the day-to-day operations from the datacenter IT team and to scale that group’s reach and efficiency. The primary responsibility of the datacenter managers should be scaling the cloud in response to growth in demand. They should be able to look at utilization in aggregate and plan and deploy space, power, networking, and servers in as much of a just in time manner as possible. They should no longer need to focus on project specific deployment activity – that is handled by a combination of the end users’ self-service activities and automated responses from the cloud itself.

Private clouds represent the transformation of an IT datacenter into a large self-service pool of resources for internal customers to use. Public clouds open up that service to external organizations to purchase. Hybrid cloud refers to when one organization uses a private cloud for some of its work and a public cloud for other work; and where there is continuity between the public and private cloud utilization. In hybrid clouds, identity, policy and user interface are common, thereby blurring the distinction between public and private to the end user. This series discusses cloud in general with public, private and hybrid being implementation decisions for specific projects. We will introduce distinctions between these cloud types only where necessary.

Cloud Business Value

The end result of cloud to the business is two-fold. Primarily, cloud enables end users to service their own IT needs in a frictionless manner, making the business much more agile – as business units and individuals can innovate quickly and execute on new ideas immediately before they become stale. A secondary but still crucial benefit is that the cost of IT and its value become intimately tied together with a high degree of transparency. Costs are minimized and associated with end user organizations and business units. Let’s dive into the cost analysis in a bit more detail.

Capex is completely eliminated for any work that can be adequately served by a public cloud. For work that needs to be done inside an organization in a private cloud, capex expenditures become just-in time and are always justified by current and projected usage. The capex, plus operating expenses, is totaled and a per unit of time per resource cost is calculated. That cost is then shown back or actually billed to the different business units. This enables the business units to make sound and informed decisions on what to deploy and not to deploy based on the value of the work they are doing. This sort of calculus is always best decided by the business unit, not IT, as they are responsible for their own P&L. In this way, capex is reduced because the business unit is incentivized to only use what they need to drive the most value to the organization.

Opex costs like power, cooling and real estate are reduced as a function of consolidating the datacenters, pooling servers, reducing wasted capacity and incentivizing end users to make lower cost requests as just described above, but they will still remain a substantial cost.

However, the administrative expense component of datacenter opex can be brought much closer to zero through infrastructure automation. The provisioning of workloads and other opex that formerly belonged to central IT is moved to the business units, not as an IT operation, but as a part of the normal flow of the day to day activities they do to get their jobs done. This relieves the business from much of its opex and/or allows the business to reallocate IT staff to more value generating activity.

Besides value to the business as a whole, cloud provides another set of benefits to the IT team specifically which should motivate IT leaders to drive the cloud discussion inside their organizations. Simply stated, cloud can make IT loved again. Historically, IT brought in technology innovations that improved the lives of their internal customers – PC’s, networks, databases, client/server applications, e-mail, mobile computing, virtualization, etc. Now, services like Amazon’s AWS have set a new expectation of IT systems responsiveness. Fairly, or unfairly, end users are coming to expect instantaneous gratification of their IT desires without the need for planning, budgeting or security audits. As a result, they are becoming impatient with IT. They wonder why IT costs so much and takes so long to deliver. By implementing a good private cloud, IT can deliver a finite set of resources in an on-demand manner without compromising security or compliance. Hybrid clouds allow IT to extend their services to handle bursts of unplanned activity that the private cloud does not have the capacity to meet. By enabling offload to public cloud while simultaneously adopting policies and controls of who can do what in public clouds, IT can introduce public cloud as another tool in the IT toolkit without abdicating their traditional responsibility for the safety of data and applications. This will prevent the skunkworks use of public clouds that too many companies see when lines of business tries to circumvent IT. Clouds can make IT the hero again.

Getting Started With Your Cloud

Now that we know what cloud is and what we should expect from it here is a proposed journey to have the easiest onramp and highest value result.

Segregate Applications

First, you need to find the right applications for cloud. As we will describe in the next article in this series, the best fits are:

  • Scale-out load balanced applications with stateless instances
  • Batch processing applications
  • Test and dev for the above two and for more traditional IT applications

As for legacy stateful IT applications deployed in production, non-cloud solutions will suffice. In many cases server virtualization will help – it can lower capex and increase service levels. Well known solutions from VMware, Microsoft, Citrix, RedHat and others can help here.

Make sure you know which end customers and which workloads are the best early cloud candidates according to the ease with which they can move to cloud and the extent to which cloud provides them with real value.

Incrementally Add Projects to the Cloud

Don’t try to boil the ocean – pick the ideal project to start with and add more challenging projects as you experience success. Plan subsequent projects incorporating knowledge gained from previous projects.

Pick a Project Based on End-User Needs and Application Type

Within the applicable projects, pick one that has a burning pain, eager customers, small enough scale to experiment, but large enough scale to be a meaningful test and that aims for results that can be measured – in terms of cost, speed to get results, lowered administration time, etc.

Pick a Private/Public/Hybrid Strategy for This Project

In a later article we will emphasize the need for your cloud software to support public, private and hybrid clouds. Assuming you have chosen a software partner that accommodates this choice, you need to make a decision for this particular project.

Choose public cloud if your project:

  • Has highly variable needs without other projects that are able to statistically offset it.
  • Is not expected to persist for an extended period of time and does not justify its own dedicated infrastructure.
  • Does not have significant needs for privacy, security, or regulatory compliance, unless there is a public cloud provider that specializes in delivering the right assurances for a project of this specific nature.

Choose private cloud if:

  • A single project or a sum of a few projects is expected to have flat or steadily growing compute needs.  Avoid projects that spike heavily and drop sharply unless they can be statistically offset by other projects that peak at different times.
  • Is it either a long-lived project or a short-lived project that, upon completion, will be replaced by projects of equal or greater resource needs.
  • Has significant needs for privacy, security, or regulatory compliance that can not be met by general-purpose or even specialty public cloud providers.

Choose hybrid cloud if the project can be meaningfully segmented into parts that have the private characteristics and other parts that have the public characteristics. Placing the right work in the right place will give you the best balance of cost, security, flexibility and return on investment.

Deploy Project

Deploying the project includes the following steps:

  • Engaging with the end user customer base to explain and sell the project.
  • Collaborate on end goals, desired metrics, and a timeframe for evaluation.
  • Acquiring, deploying and configuring physical infrastructure (if you chose private or hybrid cloud).
  • Choosing the cloud management software and deploying it.
  • Training the end-users and their management team on the self-service workflows – both for the delegation or rights and for the actual execution of work.

Evaluate

At regular points in the project, measure the results to the end user customer base and review with the customers. The goals need to be quantifiable and should be set at the beginning of the project. Goals may be around project completion time, time it takes to do specific tasks, overall cost or utilization of the infrastructure, etc. If the results are not what were expected, adjust the deployment to try to meet the goals. This can involve reconfiguring HW or SW, redoubling on customer training or adding in custom automation developed internally or from contractors.

At this point, it is also good to engage with the HW and SW providers to make sure they can identify any errors in deployment or deviation from best practices that are hampering the results. At this time, you will also have a better idea of your needs and will be in a position to make stronger and more prioritized feature requests.

Bring in Least Needful Applications for Sake of Conformity of Process, Service Levels, etc.

Only after you have had success in the ideal cloud use cases should you bring in the less applicable workloads. Though it may be harder to do so and the benefits may be less, over the long haul your IT department will benefit from a common infrastructure layer for datacenter management and a common end-user interface for deploying workloads, tracking their status and measuring their cost.

EBS, a very well done solution to one of the hardest computer science problems

Posted on by
Reply

Real cloud storage lessons from the AWS outage

It’s been very interesting watching the online firestorm over Amazon’s EBS outage.  I was not at the recent Interop show, but apparently, there was an entire panel discussion about it and then a twitter flame war between representatives from VMware and Amazon.  Then there were countless articles and blogs, all of which focused on some questions of mild interest with obvious answers.

  • Is EBS a good or bad service?
  • Will this affect people’s move to public cloud?

The answers to this are pretty uncontroversial, in my opinion. 

  • EBS is a very well done solution to one of the hardest computer science problems out there – how do you construct an infinitely scalable storage service out of commodity disk for read/write transactional data with strict consistency.  The fact that EBS has gone this long without an outage of this magnitude is a tribute to the AWS team.  They clearly made some poor choices but they will surely fix those over time.  However, few seem to be paying attention to what AWS has done correctly.  The naysayers would be hard pressed to name a production service with EBS’ characteristics deployed at EBS’ scale.
  • The best online comment on the topic of the impact to public cloud adoption likened this to an airline crash.  Airplanes crash from time to time, and those crashes always make for sensational news.  But, in terms of cost and safety, air travel remains the best way to travel long distances, so people forget the crash and keep flying.  There is a segment of people who will be nervous for some time with public cloud for some data and apps, but the trend towards public cloud will continue as before.

While all of the uproar is quite entertaining, it is not useful.  As an industry, we can be a bit more thoughtful than this.  This blog post is an attempt to get to the real lessons of this incident, which center on the industry’s transition to scale out storage models for transactional storage in the cloud.

The scale out model is undoubtedly the right architecture for cloud storage in general, especially where eventual consistency is sufficient.  It provides:

  • A highly virtualized interface – it’s one big pool of storage where placement across even thousands of nodes is completely automated
  • Great aggregate performance
  • Flexibility in the face of arbitrary failures
  • The ability to grow steadily in small increments of commodity parts as the cloud itself grows, rather than in massive chunks of proprietary equipment

But, when applied to transactional workloads, there are issues with scale out storage that (a) the vendor community still needs to work out and that (b) customers must be aware of and plan around when they take the leap.

So, why is scale-out storage for transactional workloads so hard?  To be useful to clouds, the scale-out transactional storage system needs to have the following qualities with respect to traditional enterprise storage.

  • Reliability: Cloud storage needs to be almost as reliable as enterprise storage – four or five 9’s is called for.  In this blog, I’ll speak of local availability only, not cross site-DR – that’s a topic for another day.
  • Cost: The cost of scale-out storage is expected to be considerably cheaper than enterprise storage.  Keep in mind that a lot of the enterprise storage price is in software, services, and margin.  For big cloud deals, storage vendors will negotiate down closer to the real cost of the system and/or provide leasing plans.  So, a good scale-out deployment needs to actually be cost conscious and can’t just rely on the promise of commodity parts.
  • Consistency: When using read/write transactional storage, a committed write must really be committed.  Eventual consistency does not cut it.  Any write must be truly safe as even a minute or less of data loss can be a fatal issue.
  • Performance: The transactional performance for both reads and writes must be usable, even if somewhat lower than more expensive enterprise storage systems.

With a requirement for strict consistency, protection of storage against loss and inaccessibility comes from either RAID or synchronous replication across multiple enclosures.  When you really trust an enclosure, like how people (reasonably or unreasonably) trust traditional enterprise storage, you can use RAID and minimize disk proliferation – something like 20% extra disk is a reasonable price to pay for your five 9’s.  When you don’t trust the enclosure because it’s a cheaper commodity system, you start looking at RAID over the network or erasure codes.   The challenge with this strategy is that performance can be abysmal, especially in degraded mode where each read requires too many network accesses and parity calculations.  In order to (a) make sure your data is never lost due to a critical set of commodity disks and/or enclosures being lost before rebuilds can complete while (b) maintaining adequate performance, you start mirroring the data over the network multiple times, usually 3x in scale-out systems targeted at the enterprise.  This drives up the actual cost – think TCO – disks, enclosures, power, cooling, footprint, etc…  It is notable that the mirroring system was the mechanism that brought EBS to its knees during the outage.  So, the scale-out vendor is always balancing cost, consistency, and reliability – you can get any two, but not all three at once.

Even when you achieve an acceptable balance of the first three considerations, performance can still be an issue due to economics.  To keep costs low in the face of the relatively expensive replication system, there is a temptation to pack high-density storage very tightly reducing the effective IOPs per GB and making contention a significant problem.  With all that is good about EBS, you often hear customers complaining about its performance, both in terms of maximum throughput and in terms of variability over time, even when there is no news-making outage.

All of this is just about the characteristics of the storage itself and does not deal with operational issues, which is where EBS really hit some issues.  Again, we’re not talking about TB or PB of storage, we’re talking about operations approaching the exabyte scale.

All the operations need to be completely automated – provisioning, placement, and failure response.  All this automation can be implemented in one of two places

  • Independently on each storage node
  • Through centralized controllers

The storage nodes in most scale out systems generally just store and serve data.   When data is sent to them, they store it.   When they get a read request, they serve it.  When they are given a replication partner, they send their data over.   People try to avoid putting too much logic in the storage nodes because (a) they want storage nodes to focus on streaming data and (b) if storage nodes were doing too much thinking, they’d all need to coordinate making for a potentially unsolvable peer to peer coordination problem.  Therefore, most instruction comes from one or more control systems.

In general, storage provisioning and placement operations (for primary copies, initial replicas, and new replicas after a failure), as well as data lookups, are done through more centralized controllers.  Some requirements here are as follows:

  • With thousands and thousands of users, you can’t have a single control node.  You need to have the control system itself scale-out to many, many nodes (though certainly less than the number of storage nodes). 
  • The control system needs to be even more accessible than the data.  You never want a situation where the control service nodes all die, get confused, lose metadata, or are simply starved for resources as when that happens, all administrators and end users lose the ability to interact with storage system as a whole.
  • The algorithms of the control system need to be very clever since they are controlling thousands and thousands of individual storage nodes, which generally obey even inappropriate and/or heavyweight commands faithfully. 

If you read Amazon’s well written, open, and frank EBS post-mortem, you know how and where these guidelines were violated and where the EBS team will undoubtedly be placing their efforts to improve the service over time.  But, for you, the cloud builder, here is what you need to talk about when you talk to a scale-out storage provider.

1) What tradeoffs were made between reliability, cost, and consistency?   If you need strong consistency on transactional data, find out what the uptime guarantees are, and what the implications are for overall system cost.  Dig deep into any uptime guarantees.  Make sure you understand the assumptions regarding probability of individual failures and adjust those assumptions if they do not apply exactly in your datacenter.

2) What is the price per usable GB and per IOP?   If you are building a big enough cloud and can negotiate a great price from an enterprise storage vendor, make sure the scale-out system is cost-competitive even though they will be using many more disks.   Think about TCO – don’t forget about the power, cooling, and footprint costs that come along!  This is not to say that scale-out is more expensive than traditional storage, or that you should not go for it if you don’t get the savings you hope for.  But you should double check all the math and make sure the TCO is what you expect.

3) What is the performance of the transactional storage– both in a normal mode and in a degraded mode?   Make sure that they are not assuming a lower spindle to IOPs ratio than is reasonable (like EBS) to give you a rosy picture on price.  Assume your transactional storage will actually be accessed forcing you to increase spindles, use less dense storage, and/or have a really good caching/tiering/ILM story.

4) What are the assumptions of the storage system?   The EBS design assumed that their redundant network would always be available and that there would never be a general loss of connectivity from all to all.  Has your scale-out vendor designed for this eventuality and tested it at scale?  What other datacenter assumptions are they making?

5) What happens with split-brain at scale?  Traditional enterprise storage is very simple in this area.  Local availability is handled inside a single chassis and DR is done with dedicated replication partnerships.   It is inflexible and not responsive to changing conditions.  Scale-out storage is way better in this regard, but if not done right, the flexibility of the scale-out system can backfire, just like in the EBS case where all nodes tried to re-establish replicas of all data at once.

6) Does the storage understand temporary vs. permanent outages?  If so, what if something that appeared permanent turns out to be actually temporary?   Can your storage system react to the return of service in a reasonable way, especially when the permanent failure response is very heavyweight?  EBS, unlike traditional enterprise storage, kept re-mirroring to new nodes rather than simply sync’ing back up with old mirrors when they once again became accessible.

7) Can the control system guarantee access to users and administrators?  In the EBS outage, the automated failure response overloaded the control service, which is what actually affected all users, even if they had properly replicated their data between availability zones.

8) Are your availability zones really isolated?  In EBS, there was a shared resource between availability zones.   This is what made the impact of the response to #7 so bad.

9) Does the automation know when to stop trying something?   Once it was clear that no more space was to be had and that the control systems were not responsive, the automated re-protection kept going.  Sometimes, like people, software needs to stop, take a breath and let the situation cool down.  And even though this is cloud, when the storage is in this state, it’s best to have it ask for administrator intervention rather than continuing to try to do the impossible repeatedly.

10) Are failures graceful, even the unlikely ones?   The EBS system had a corner case that crashed the nodes rather than failing an operation gracefully.  In most software, you can get away with letting those corner cases go, but when approaching the exabyte scale, you can’t.  Make sure your vendor has good software engineering practices here.

11) Are there good fail-safes?   The EBS outage started to get better when the EBS admins were able to stop some of the communication and get out of the vicious cycle.  Does your scale-out vendor have similar controls to allow you to manually stop heavyweight operations that you, as the cloud operator, determine need stopping for the sake of the cloud as a whole?

12) Are the requirements for the end customer documented?  After the outage, Amazon put out some excellent documentation on building cloud applications that everyone should read.  Does your scale-out system, due to performance or reliability tradeoffs, require end users to use the storage system in any specific and non-obvious ways?  If so, make sure those are clearly documented so you can educate your end users.

While items 4-10 in this list derive from the EBS problems, this blog posting should not be seen as anti-EBS.   With EBS, Amazon has created something unique in the industry, a massive read/write transactional storage system with strong consistency that can be operated by a reasonable sized IT staff.   Its major outage was the first of this level of seriousness in years and the long-term affects have been quite minimal.   The success of EBS has influenced the rise of a plethora of scale out storage startups that want to give you something EBS-like in your datacenter.   It has also scared the traditional storage vendors on technology and pricing and pushed them to innovate in a way they’ve not done in a long time – see their recent product announcements and M&A activity.  EBS is a great service that will only get better. 

While EBS’ failure in this case was spectacular, in a way it was fortuitous for the cloud industry because it educates us on what to look for in storage vendors.  Hopefully, the scale-out storage vendors have been paying attention as well.  They can learn important lessons about operations at massive scale without needing to do a very expensive real-world QA and without causing an outage for a paying customer.   These lessons should be the focus of our attention, not the drama.