Popcorn

Abstract

The recent possibility of integrating multiple-OS-capable, high-core-count, heterogeneous-ISA processors in the same platform poses a question: given the tight integration between system components, can a shared memory programming model be adopted, enhancing programmability? If this can be done, an enormous amount of existing code written for shared memory architectures would not have to be rewritten to use a new programming paradigm (e.g., code offloading) that is often very expensive and error prone. We propose a new software architecture that is composed of an operating system and a compiler framework to run ordinary shared memory applications, written for homogeneous machines, on OS-capable heterogeneous-ISA machines. Applications run transparently amongst different ISA processors while exploiting the most optimized instruction set for each code block. We have implemented and tested our system, called Popcorn, on a multi-core Intel Xeon machine with a PCIe Intel Xeon Phi to demonstrate the viability of our approach. Application execution on Popcorn demonstrates to be up to 52% faster than the most performant native execution on Linux, on either Xeon or Xeon Phi, while removing the burden of the programmer having to adopt a different programming model than shared memory on a heterogeneous system. When compared to an offloading programming model, Popcorn is shown to be up to 6.2 times faster.

libMPNode

“…Traditionally, developers have used the message passing interface (MPI) to distribute execution across nodes [11]. Deemed the "assembly language of parallel processing" [18], MPI forces developers to orchestrate parallel computation and manually keep memory consistent across nodes through low-level send/receive APIs, which leads to complex applications [4]. Partitioned global address space (PGAS) languages like Unified Parallel C [9] and X10 [7] provide language, compiler and runtime features for a shared memory-esque abstraction on clusters.…”

“…Because the DSM is implemented transparently by the OS, existing sharedmemory applications can execute across nodes unmodified. The complete details of Popcorn Linux can be found in past works [3,4,15,16,25].…”

“…Because of these trends and due to the complexities of scaling interconnects to larger core counts [17], the systems community has experienced new interest in moving from software architectures for fully cache-coherent systems to those composed of multiple incoherent domains, where cache coherence is provided only within a domain, as shown in Figure 1. Recent works have developed software architectures for mobile SoCs containing incoherent CPU cores [1,19], servers with incoherent compute elements [3,4,12] or even rack-scale clusters [21][22][23]. In these architectures, the software treats each cache coherence domain as a node and provides cross-node execution and memory coherence transparently to applications.…”

Rebooting Virtualization Research (Again)

“…Obviously, the hypervisor of a disaggregated rack should be distributed, meaning that each resource blade runs a piece of the hypervisor. Building on existing OS paradigms, two main approaches can be envisioned for the construction of such a distributed hypervisor, namely multi-hypervisor (analog to multi-kernel [2,3]) and split-hypervisor (analog to split-kernel [30]). In the former, the hypervisor is a collection of several complete hypervisors.…”

“…The question is how to guarantee certain performance levels for this path. Several solutions can be adopted including: (i) building CPU blades with high memory cache sizes (gigabytes) and optimizing the hypervisor to efficiently exploit this local cache [29,30]; (2) building dedicated network links between CPU and memory blades; (3) allowing bandwidth reservations on a shared interconnect.…”