Skip to: Site menu | Main content

Anastassios Nanos

Researcher at CSLab, NTUA

~$ cat /dev/research

Major interests

V4VSockets

In the HPC context, applications often scale to a large number of nodes, leading to the need for a high-performance interconnect to provide low-latency and high-bandwidth communication. In the cloud context, distributed applications are executed in a set of Virtual Machines (VMs) that are placed in physical nodes across the cloud data center. These VMs suffer communication overheads, due to various intermediate layers that abstract away the physical characteristics of the underlying hardware and multiplex the application’s access to I/O resources. Moreover, these VMs are unaware of their physical placement – this presents a problem, because application instances running on the same physical node but on different VMs are not able to exploit locality. Data that need to be exchanged between instances of the same application reside in the physical memory of the node. However, VMs are not aware of this and as a result, data flow through various unneeded layers (the network stack, etc.) until the actual message exchange is realized.

We develop V4VSockets, a socket-compliant, high-performance intra-node communication framework for co-located Virtual Machines. V4VSockets contains many features that reduce and eliminate problems associated with traditional PV drivers in an HPC context. Specifically, it simplifies the data path between co-located VMs. This is achieved by creating a peer-to-peer communication channel between a VM and the hypervisor. V4VSockets eliminates the overhead of page exchange/mapping and enhances throughput by moving the actual copy operation to the receiver VM. V4VSockets improves security by operating in a shared-nothing policy; pages are not shared between VMs – the hypervisor is the only one responsible for transferring data to the peer VM. Moreover, V4VSockets complies to the classic network layer concept, thus simplifying the interface to applications.

Preliminary code is available online (v4vsockets).

Xen2MX

As we move towards the standardization of Ethernet in both worlds, Cloud computing and High-performance computing, we need a way to to study the effect of message-passing protocols in the Cloud, without having to suffer TCP/IP’s complexity. However, current approaches do not provide a software solution to efficiently exploit hypervisor abstractions to access hardware. We move forward to a more generic design, in order to understand and optimize the way VMs communicate with the network in an HPC context.

We design Xen2MX, a high-performance interconnection protocol for virtualized environments. Xen2MX is binary compatible with MX and wire compatible to MXoE, the ethernet mode of Myrinet’s MX protocol. Although our prototype implementation is in early stages, results from the original MX benchmarks over Xen2MX are promising: virtualization overheads are almost eliminated compared to a software bridge setup, the generic way of communication in virtualized environments. We are in the process of finalizing our implementation and examine possible ways to optimize our prototype. The code is available on github and on a local gitorious installation.

WiP

We believe that modern High Performance Interconnection Networks provide abstractions that can be exploited in Virtual Machine execution environments but lack support in sharing architectures. Previous work has shown that integrating the semantics of Virtualization in specialized software that runs on Network Processors can isolate and finally minimize the overhead on the VM Hypervisor concerning access to the device by Guest VMs. Direct I/O has been proposed as the solution to the CPU overhead imposed by guest VM transparent services that can lead to low throughput for high bandwidth links. However, minimizing the CPU overhead comes at the cost of giving away the benefits of the device driver model. Integrating protocol offload support (present in most modern NICs) in virtual network device drivers can lead to performance improvement. Bypassing the Hypervisor while moving data around, can also minimize the overhead imposed by heavy I/O but at the cost of security and isolation.

We envision a Virtualization-enabled High performance Network Interface that can achieve line-rate throughput and optimized sharing of Network I/O in Virtual Machines by utilizing commodity hardware and innovative resource-sharing virtualization architectures.

Data access in HPC infrastructures is realized via user-level networking and OS-bypass techniques through which nodes can communicate with high bandwidth and low-latency. Virtualizing physical components requires hardware-aided software hypervisors to control I/O device access. As a result, line-rate bandwidth or lower latency message exchange over 10GbE interconnects hosted in Cloud Computing infrastructures can only be achieved by alleviating software overheads imposed by the Virtualization abstraction layers, namely the VMM and the driver domains which hold direct access to I/O devices. We have designed MyriXen, a framework in which Virtual Machines efficiently share network I/O devices bypassing overheads imposed by the VMM or the driver domains. MyriXen permits VMs to optimally exchange messages with the network via a high performance NIC, leaving security and isolation issues to the Virtualization layers. Smart Myri-10G NICs provide hardware abstractions that facilitate the integration of the MX semantics in the Xen split driver model. With MyriXen, multiple VMs exchange messages using the MX message passing protocol over Myri-10G interfaces as if the NIC was assigned solely to them. We believe that MyriXen can integrate message passing based application in clusters of VMs provided by Cloud Computing infrastructures with near-native performance.