This days Microsoft launched two new types of instances – A8 and A9. This two new type of instances are like a bulldozer from the perspective of the available RAM memory and processor. Each of this machines has 8 and 16 virtual cores and 56GB and 112GB of RAM memory. Yes, you heard me good 112GB on RAM memory available on cloud instances.
For customers, that has intensive work to do this is a great news. Starting from now it will be pretty easy to have an intensive processor tasks that are processed on the cloud, without having to use multiple machines. Another opportunity is for the applications that needs to use a lot of memory – cache systems for example. For this kind of system, having more than 100GB of memory available is great. If we think a little bit, we can imagine a scenario where we have 100GB of data loaded from a relational database into a cache. This is great!
But, this is not all. This new kind of instances are more than an instance of steroids. This are a new generation of instances, that has not only more CPU and RAM but also other new features (see below).
Having an instance on steroids can be without real value if you don’t provide the perfect environment where to use all this power. Because of this, A8 and A9 instances have Remote Direct Memory Access (RDMA) and InfiniBand.
RDMA is a features that allows two different computer to access directly the same memory without using the operation system infrastructure. Basically, in this way we have a low-latency network with an extremity high-throughput. This is a part of the key to have an instance that can be used for massively parallel processing.
This mechanism allow us to transfer data from application with zero-copy time (we don’t need to copy the content between application or use buffers). The advantage of this feature is almost 0 resources required from CPU, cache or context switches. Theoretically, using this feature we can move 4GB of data that is in memory between two machines in under 1 second.
The second feature, which unlocked RDMA and also allowed high throughput is InfiniBand. This allow a fast connection between processor nodes and I/O nodes. In this way all the resources can be used at maximum level. If you hear about InfinitiBand for the first time, you should know that this is not a traditional Ethernet network architecture. It is based on a switched fabric topology. All the information that is transmitted is in form of packages that form a message. Each message can have different types like:
This two feature enabled Messaging Passing Interface (MPI). This system is to sustain application that needs to be large scalable application. You should know that MPI is not specific to Microsoft stack (you can use different languages to write application for it Fortran).
This new types of instances can be used with success for application from different domain likes: weather modeling, science, different simulation or intensive computing tasks that can be parallelized.
For customers, that has intensive work to do this is a great news. Starting from now it will be pretty easy to have an intensive processor tasks that are processed on the cloud, without having to use multiple machines. Another opportunity is for the applications that needs to use a lot of memory – cache systems for example. For this kind of system, having more than 100GB of memory available is great. If we think a little bit, we can imagine a scenario where we have 100GB of data loaded from a relational database into a cache. This is great!
But, this is not all. This new kind of instances are more than an instance of steroids. This are a new generation of instances, that has not only more CPU and RAM but also other new features (see below).
Having an instance on steroids can be without real value if you don’t provide the perfect environment where to use all this power. Because of this, A8 and A9 instances have Remote Direct Memory Access (RDMA) and InfiniBand.
RDMA is a features that allows two different computer to access directly the same memory without using the operation system infrastructure. Basically, in this way we have a low-latency network with an extremity high-throughput. This is a part of the key to have an instance that can be used for massively parallel processing.
This mechanism allow us to transfer data from application with zero-copy time (we don’t need to copy the content between application or use buffers). The advantage of this feature is almost 0 resources required from CPU, cache or context switches. Theoretically, using this feature we can move 4GB of data that is in memory between two machines in under 1 second.
The second feature, which unlocked RDMA and also allowed high throughput is InfiniBand. This allow a fast connection between processor nodes and I/O nodes. In this way all the resources can be used at maximum level. If you hear about InfinitiBand for the first time, you should know that this is not a traditional Ethernet network architecture. It is based on a switched fabric topology. All the information that is transmitted is in form of packages that form a message. Each message can have different types like:
- Atomic Operation
- Multicast Transmission
- Channel Send or Receive
- Transaction Based Operation
- Direct Memory Access (RDMA)
This two feature enabled Messaging Passing Interface (MPI). This system is to sustain application that needs to be large scalable application. You should know that MPI is not specific to Microsoft stack (you can use different languages to write application for it Fortran).
This new types of instances can be used with success for application from different domain likes: weather modeling, science, different simulation or intensive computing tasks that can be parallelized.
Comments
Post a Comment