HMTT：硬件/软件追踪系统：弥合DRAM访问跟踪的语义差距

167 浏览量更新于2024-07-14 1 收藏 1.37MB PDF 举报

HMTT（Hybrid Hardware/Software Tracing System）是一项旨在解决DRAM访问跟踪语义鸿沟的研究论文。在现代计算机系统中，DRAM访问跟踪（即芯片外内存引用）对于内存子系统的优化和软件性能调优至关重要。硬件嗅探，作为一种在芯片外内存接口上监测和收集实际DRAM访问的非侵入式方法，具有高效的优势。然而，与基于软件的方法相比，硬件嗅探通常缺乏关键的语义信息，如进程/函数标识符、虚拟地址以及锁定上下文，这些信息对于深入理解系统和软件行为是必不可少的。本文的主要贡献在于提出了一种混合硬件和软件的追踪系统，HMTT。这种设计的目标是弥补现有硬件嗅探方法的局限性，通过结合硬件和软件的优势来提供更全面的系统和软件行为理解。具体来说，HMTT可能包括以下几个关键技术点： 1. **硬件辅助数据采集**：硬件层面可能实现高效的实时内存访问监控，捕获到大量的底层物理地址和时间戳等基础信息。 2. **中间件解析**：将这些原始硬件数据转换成包含丰富语义的事件流，例如通过软件解码，解析出虚拟地址映射、进程ID等。 3. **跨层通信**：硬件和软件组件之间可能通过一种高效、低延迟的接口进行协作，以便实时同步和共享信息。 4. **数据存储和处理**：软件部分负责对处理过的数据进行聚合、分析和存储，支持进一步的性能分析或故障诊断。 5. **用户界面和可视化**：为了便于用户理解和利用这些信息，系统可能提供直观的用户界面，展示内存访问模式、性能瓶颈和潜在优化建议。 6. **可扩展性和灵活性**：HMTT的设计应考虑未来技术演进，允许适应不同的硬件架构和软件环境。通过HMTT，研究人员和开发者能够获得更详尽的内存访问洞察，从而提升系统性能、优化资源分配，或者更好地调试和分析软件问题。该工作不仅对硬件设计有直接影响，也对提高整体系统的理解和优化提供了新的可能性。

HMTT: Bridging the DRAM Access Trace’s Semantic Gap 7:5

Fig. 2. Example of a hybrid hardware/software tracing mechanism.

investigation. However, the high-level events are usually insufﬁcient for studying the

performance and behavior of a system in depth.

Based on these observations, we can conclude that there is a semantic gap between

conventional DRAM access traces and high-level events. If they can be correlated with

one another, as shown in Figure 1(c), it should be signiﬁcantly helpful for both low-level

(DRAM access traces) and high-level (system or program events) analysis. Although

a similar gap exists between cache requests and high-level events, many mechanisms

such as Intel’s PEBS [Intel Corporation 2012] can solve i t by directly locating the

sources of events for cache requests.

However, the current trace tools can only collect either DRAM access traces or high-

level events such as function call graphs and OS events. Some hardware monitors

are only capable of collecting whole-memory requests by snooping on the memory

bus; examples are those described by Alexander et al. [1986] and Fuentes [1993], and

Monster [Nagle et al. 1992], MemorIES [Nanda et al. 2000], and ACE [Hong et al. 2006].

For high-level events, gprof can provide only call graphs, and the Linux Trace Toolkit

[Desnoyers and Dagenais 2006] focuses on collecting OS events; however, these have a

substantial amount of overhead. In addition, by instrumenting the target program with

additional instructions, some instrumentation tools such as ATOM [Srivastava and

Eustace 1994], Pin [Luk et al. 2005], and Valgrind [Nethercote and Seward 2007] are

capable of collecting more information, for example, memory traces and function call

graphs. However, such instrumentation may cause extra overhead, disturb program

execution, and cause distortion problems. As a result, the information collected by

such instrumentation tools may be inaccurate or even useless. Moreover, it is hard to

instrument virtual machine monitors and operating systems.

In summary, there is a semantic gap between conventional DRAM access traces and

high-level events, but almost none of the existing tools are capable of bridging the gap

effectively.

3. A HYBRID HARDWARE/SOFTWARE TRACING MECHANISM

To address this semantic gap, we propose a hybrid hardware/software mechanism

that is able t o collect memory reference traces and high-level event information

simultaneously.

Figure 2 shows an overview of the use of a hybrid hardware/software mechanism to

study the pattern of object accesses. A semantic gap exists here between the DRAM

access trace and the object identiﬁers. On the one hand, low-overhead hardware is used

ACM Transactions on Architecture and Code Optimization, Vol. 11, No. 1, Article 7, Publication date: February 2014.

剩余24页未读，继续阅读

weixin_38691669

粉丝: 3

HMTT：硬件/软件追踪系统：弥合DRAM访问跟踪的语义差距

hmtt:在里面

一种软硬件结合的大数据访存踪迹收集分析工具集.docx

hmtt-multiport：头条多端案例

centos7hmtt

基于Vue框架的hmtt项目源码，探讨尤玉溪对Vue的不懂之处

C语言基础学习常用数据结构和算法

Short Text Hashing Improved by Integrating Multi-Granularity Topics and Tags

system:后台管理系统

cole_02_0507.pdf

工程硕士开题报告：无线传感器网络路由技术及能量优化LEACH协议研究

最新资源