�ݺ�ߣ

Editor's Notes

1. Challenges for multi-tenancy include optimisation of resource sharing and guaranteed isolation against physical limitations, co-located faults and malicious attacks. For these reasons testing the performance and resilience of applications with different hardware, platforms, configurations, resource sharing conditions and user loads is important for increasing the assurance that providers and consumers have in cloud applications.
2. Scalability: given that customer applications can vary from 1 to 1000s of nodes, it must be possible to readily set up and execute useful testing environments for 1 to n number of independent hosts and network elements, avoiding cumbersome, error-prone configuration.Reproducibility: it must be possible to easily repeat testing conditions and tests in order to perform viable regression testing and make reliable claims about software quality.Portability: as hardware and virtualization technologies change, or as applications may be migrated to alternative data centers and platforms, it should be possible to easily reuse and recreate test designs and conditions during these changes.Extensibility: in addition to portability, test designs will need to be modified over time in order to take into account changes in quality constraints, scope, expected demands and software functionality. Systematic, guided procedures for modifying and extending test designs and mechanisms for these changes are necessary, as opposed to starting from scratch each time or making changes without knowledge of all dependencies.Self-containment: it is desirable to have a single top level solution, operational interface and workflows for designing and executing tests as opposed to the tester having to switch between multiple contexts and tools.Controllability: in spite of abstraction and higher level tooling, there is still a need to have control over the behaviour of resources used in the test, minimising the amount of disturbances that might provide unknown variations in test results.
3. Each requirement is assessed considering four different perspectives that arise in real-world systems with changing business priorities and technologies:Different test types (e.g. functional, load, security) need to be performed.A variety of resource kinds (i.e. network, disk, CPU, memory) need to be manipulated.Heterogeneity of physical hosts, nodes and devices.Given different customers, different scenarios and workload mixes have to be considered in parallel.
4. The main principles behind the design of the framework are the abstraction from the lower level tools that are used to create loads over the resources, allowing to emulate the wanted contention, as well as the separation of concerns, providing an effective modularisation of the tool which enables easy extensibility and additions of emulators and distributions.
5. A distribution creates the trend through runs as in a sampling process
6. Given a share physical machine, we divide the physical resources to VMs (SuT and COCOMA) each time essentially with a different percentage of the total.
7. CPU, the SuT CPU benchmark degradation over the percentage of CPU used by COCOMAThree clusters can be identifiedin the highest one there are configurations 1, 4 and 5 which according to table II have all the MH setup for the CPU for COCOMAbelow there are configurations 2, 7 and 8, which have the ML setup for the CPU for COCOMAand in the lowest part there are 6 and 3, with COCOMA CPU setup as LAs expected the more physical CPU is controlled by COCOMA the more the SuT is affected when CPU intensive operations are performedRAM, COCOMA used the maximum assigned RAM while increasing the number of threads performing writing operations on the RAMAmount of RAM and overall number of VMs assigned to COCOMA influence the resultsconfigurations 1, 2, 3 and 6, which have all 2 VMs in total, the more amount of RAM is assigned to COCOMA the more the SuT is affectedif we compare configuration 5 and 1, the difference in results is due only to the total number of VMs, being the only differentiator parameter between the two configurationsIOthe amount of files used for the workload does not make any noticeable difference across all configurationsConfiguration 4 and 5 (with 3 VMs assigned to COCOMA vs 1 to the SuT) suffers most
8. Performance testers/engineers: practitioners investigating for example new colocation algorithms, and generally in need to create a contentious/malicious environment to conduct their tests;Cloud Service Providers: in this case a service provider may offer performance isolation mechanisms and therefore wants to test the effectiveness of those mechanisms, investigate the possibility of offering those mechanisms, or study what characteristics applications need to coexist;Cloud Application Administrators: administrators may need to check when a system is restored, after some maintenance or a crash, that performance isolation mechanisms are working correctly;Application Developers and Testers: application developers may want to investigate the effects of contention over their system, while testers may want to check if provider’s isolation mechanisms work sufficiently;Benchmarks and Standards Groups: in this case it can be used to validate cloud patterns and workloads under investigation and/or characterisation.
9. On the malicious part, we are looking into covert channels (or side channels) at cache level to infer other processes information and data, as well as at network level to get information about co-located guests, such as IP addresses. The latter could be used by fuzzers to send malicious workloads, or do a DoS.
10. BonFIRE is an EU project which is designing, building and operating a multi-site cloud-based facility on top of six infrastructure offering heterogeneous Cloud resources, including compute, storage and network. BonFIRE is geared towards experimentation and research into Cloud/IoS, and offers the facilities to easily create, manage and monitor experiments, whilst giving the experimenters more information and control of Cloud resources than what is offered by other public Cloud providers.