`Published as Operating Systems Review 34(5):202–216, Dec. 1999
`
`Design and implementation of a distributed virtual
`machine for networked computers
`
`Emin Gün Sirer, Robert Grimm, Arthur J. Gregory, Brian N. Bershad
`
`University of Washington
`Department of Computer Science and Engineering
`
`{egs, rgrimm, artjg, bershad}@cs.washington.edu
`
`Abstract
`This paper describes the motivation, architecture and
`performance of a distributed virtual machine (DVM) for
`networked computers. DVMs rely on a distributed service
`architecture to meet the manageability, security and
`uniformity requirements of large, heterogeneous clusters of
`networked computers. In a DVM, system services, such as
`verification,
`security enforcement, compilation and
`optimization, are factored out of clients and located on
`powerful network servers. This partitioning of system
`functionality reduces resource requirements on network
`clients, improves site security through physical isolation
`and
`increases
`the manageability of a
`large and
`heterogeneous network without sacrificing performance.
`Our DVM implements the Java virtual machine, runs on
`x86 and DEC Alpha processors and supports existing Java-
`enabled clients.
`
`1. Introduction
`Virtual machines (VMs) have the potential to play an
`important
`role
`in
`tomorrow’s networked computing
`environments. Current trends indicate that future networks
`will likely be characterized by mobile code [Thorn 97],
`large numbers of networked hosts per domain [ISC 99] and
`large numbers of devices per user that span different
`hardware architectures and operating systems [Hennessy
`99, Weiser 93]. A new class of virtual machines,
`exemplified by systems such as Java and Inferno [Lindholm
`& Yellin 96, Dorward et al. 97], has recently emerged to
`meet the needs of such an environment. These modern
`virtual machines are compelling because they provide a
`
`Permission to make digital or hard copies of all or part of this
`work for personal or classroom use is granted without fee
`provided that copies are not made or distributed for profit or
`commercial advantage, and that copies bear this notice and the
`full citation on the first page. To copy otherwise, to republish, to
`post on servers or to redistribute to lists, requires prior specific
`permission and/or a fee.
`SOSP-17 12/1999 Kiawah Island, SC
`© 1999 ACM 1-58113-140-2/99/0012…$5.00
`
`platform-independent binary format, a strong type-safety
`guarantee that facilitates the safe execution of untrusted
`code and an extensive set of programming interfaces that
`subsume those of a general-purpose operating system. The
`ability to dynamically load and safely execute untrusted
`code has already made the Java virtual machine a
`ubiquitous component in extensible systems ranging from
`web browsers and servers to database engines and office
`applications. The platform independence of modern virtual
`machines makes it feasible to run the same applications on
`a wide range of computing devices, including embedded
`systems, handheld organizers, conventional desktop
`platforms and high-end enterprise servers. In addition, a
`single execution platform offers the potential for unified
`management services, thereby enabling a small staff of
`system administrators to effectively administer thousands or
`even hundreds of thousands of devices.
`While modern virtual machines offer a promising
`future, the present is somewhat grim. For example, the Java
`virtual machine, despite
`its commercial success and
`ubiquity, exhibits major shortcomings. First, even though
`the Java virtual machine was explicitly designed for
`handheld devices and embedded systems, it has not been
`widely adopted in this domain due to its excessive
`processing and memory requirements [Webb 99]. Second, it
`is the exception, rather than the rule, to find a secure and
`reliable Java virtual machine [Dean et al. 97]. And third,
`rather than simplifying system administration, modern
`virtual machines, like Java, have created a substantial
`management problem [McGraw & Felten 96], leading many
`organizations to simply ban virtual machines altogether
`[CERT 96].
`We assert that these symptoms are the result of a much
`larger problem that is inherent in the design of modern
`virtual machines. Specifically, state of the art modern
`virtual machines rely on the monolithic architecture of their
`ancestors [Goldberg 73, Popek & Goldberg 74, IBMVM
`86, UCI 96]. All service components in a monolithic VM,
`such as verification, security management, compilation and
`optimization, reside locally on the host intended to run the
`
`202
`
`PALO ALTO NETWORKS Exhibit 1004 Page 1
`
`
`
`VM applications. Such a monolithic service architecture
`exhibits four shortcomings:
`1. Manageability: Since each modern virtual machine
`is a completely independent entity, there is no
`central point of control
`in an organization.
`Transparent and comprehensive methods
`for
`distributing security upgrades, capturing audit trails
`and pruning a network of rogue applications are
`difficult to implement.
`2. Performance: Modern virtual machine services,
`such as authentication, just-in-time compilation and
`verification, have substantial processing and
`memory requirements. Consequently, monolithic
`systems are not suitable for hosts, such as
`embedded devices, which lack the resources to
`support a complete virtual machine.
`3. Security: The trusted computing base (TCB) of
`modern VMs
`is not small, well-defined, or
`physically isolated from application code. A large
`TCB with ill-defined boundaries makes it difficult
`to construct and certify secure systems [Saltzer &
`Shroeder 75]. The lack of separation between
`virtual machine components means that a flaw in
`any component of the virtual machine can place the
`entire machine at risk [McGraw & Felten 99].
`Further, co-location of VM services has resulted in
`non-modular systems that can exhibit complex
`inter-component
`interactions, as observed for
`monolithic operating systems [Accetta et al. 86,
`Bershad et al. 95, Engler et al. 95].
`4. Scalability: Monolithic virtual machines are
`difficult to port across the diverse architectures and
`platforms found in a typical network [Seltzer 98].
`In addition, they have had problems scaling over
`the different usage requirements encountered in
`organizations [Rayside et al. 98].
`The goal of our research is to develop a virtual machine
`system that addresses the manageability, performance,
`security and
`scalability
`requirements of networked
`computing. In addition, such a system should preserve
`compatibility with the wide base of existing monolithic
`virtual machines in order to facilitate deployment. To this
`end, we focus on implementation techniques that preserve
`the external interfaces [Lindholm & Yellin 96] and platform
`APIs [Gosling & Yellin 96] of existing virtual machines.
`We address
`the problems of monolithic virtual
`machines with a novel distributed virtual machine
`architecture based on service factoring and distribution. A
`distributed service architecture factors virtual machine
`services into logical components, moves these services out
`of clients and distributes them throughout the network. We
`have designed and implemented a distributed virtual
`machine for Java based on this architecture. Our DVM
`
`includes a Java runtime, a verifier, an optimizer, a
`performance monitoring service and a security manager. It
`differs from existing systems in that these services are
`factored into well-defined components and centralized
`where necessary.
`The rest of the paper is structured as follows. The next
`section describes our architecture and provides an overview
`of our system. Section 3 describes the implementation of
`conventional virtual machine
`services under our
`architecture. Section 4 presents an evaluation of the
`architecture and Section 5 shows how a new optimization
`service can be accommodated under this architecture.
`Section 6 discusses related work; Section 7 concludes.
`2. Architecture overview
`The principal insight behind our work is that centralized
`services simplify service management by reducing the
`number and geographic distribution of the interfaces that
`must be accessed in order to manage the services. As
`illustrated by the widespread deployment of firewalls in the
`last decade [Mogul 89, Cheswick & Bellowin 94], it is far
`easier to manage a single, well-placed host in the network
`than to manage every client. Analogously, we break
`monolithic virtual machines up into their logical service
`components and factor these components out of clients into
`network servers.
`for a virtual machine
`The service architecture
`determines where, when and how services are performed.
`The location (i.e. where), the invocation time (i.e. when),
`and
`the
`implementation (i.e. how) of services are
`constrained by the manageability, integrity and performance
`requirements of the overall system, and intrinsically involve
`engineering
`tradeoffs. Monolithic virtual machines
`represent a particular design point where all services are
`located on the clients and most service functionality,
`including on the fly compilation and security checking, is
`performed during the run-time of applications. While this
`paper shows the advantages of locating services within the
`network, changing the location of services without regard
`for
`their
`implementation can significantly decrease
`performance as well. For instance, a simple approach to
`service distribution, where services are decomposed along
`existing interfaces and moved, intact, to remote hosts, is
`likely to be prohibitively expensive due to the cost of
`remote communication over potentially slow links and the
`frequency of inter-component interactions in monolithic
`virtual machines. We describe an alternative design where
`service
`functionality
`is
`factored out of clients by
`partitioning services into static and dynamic components
`and present an implementation strategy that achieves
`performance comparable to monolithic virtual machines.
`In our distributed virtual machine, services reside on
`centralized servers and perform most of their functionality
`statically, before the application is executed. Static service
`
`203
`
`PALO ALTO NETWORKS Exhibit 1004 Page 2
`
`
`
`Static Service Components
`
`Perimeter Services
`
`Verifier
`
`Security
`Enforcement
`
`Management Svcs
`Client
`Manager
`
`Execution Svcs
`
`Compiler
`
`Optimizer
`
`Cache
`
`Internet
`
`Auditer
`
`Profiler
`
`Clients
`
`Runtime
`
`Runtime
`Runtime
`
`Dynamic Service
`Components
`
`Network
`Management
`Server
`Security
`Server
`Library
`Manager
`
`Administration
`Console
`
`Figure 1. The organization of static and dynamic service components in a distributed virtual machine.
`
`components, such as a verifier, compiler, auditor, profiler,
`and optimizer, examine
`the
`instruction segment of
`applications prior to execution to ensure that the application
`exhibits the desired service properties. For example, a
`verifier may check the code for type-safety, a security
`service may examine the statically determinable arguments
`to system calls, and an optimizer may check code structure
`for good performance along a particular path.
`The dynamic service components provide service
`functionality during the execution of applications. They
`complement static service components by providing the
`services that inherently need to be executed at application
`run-time in the context of a specific client. For example, a
`security service may check user-supplied arguments to
`system calls, a profiler may collect run time statistics, and
`an auditing service may generate audit events based on the
`execution of the application.
`The glue that ties the static and dynamic service
`components together is binary rewriting. When static
`service components encounter data-dependent operations
`that cannot be performed statically, they insert calls to the
`corresponding dynamic service components. For example,
`our static verification service checks applications for
`conformance against the Java VM specification. Where
`static checking cannot completely ascertain the safety of the
`program, the static verifier modifies the application so that
`it performs the requisite checks during its execution. The
`
`resulting application is consequently self-verifying because
`the checks embedded by the static service component are an
`integral part of the application code.
`Figure 1 illustrates our distributed virtual machine
`architecture. Static service components produce self-
`servicing applications, which require minimal functionality
`on the clients. Dynamic service components provide service
`functionality to clients during run-time as necessary. The
`static services in our architecture are arranged in a virtual
`pipeline that operates on application code, as shown in
`Figure 2.
`A distributed service architecture allows the bulk of VM
`service functionality to be placed where it is most
`convenient. A natural service placement strategy is to
`structure the static service components as a transparent
`network proxy, running on a physically secure host. Placed
`at a network trust boundary, like a firewall, such a proxy
`can transparently perform code transformations on all code
`that
`is
`introduced
`into an organization.
`In some
`environments, the integrity of the transformed applications
`cannot be guaranteed between the server and the clients, or
`users may introduce code into the network that has not been
`processed by the static services. In such environments,
`digital signatures attached by the static service components
`can ensure that the checks are inseparable from applications
`[Rivest et al. 78, Rivest 92], and clients can be instructed to
`redirect
`incorrectly signed or unsigned code
`to
`the
`
`Code
`
`Cache
`Check
`
`Verifier Security Compiler Optimizer Profiler Cache
`
`Runtime
`
`check
`type
`safety
`
`translate
`to native
`format
`
`check
`static
`rules
`annotate
`for
`dynamic
`checks
`Figure 2. The flow of code through a pipeline of static service components in a distributed virtual machine. The ordering of
`services in this pipeline may be modified to suit organizational or functional requirements. Further, the client runtime may
`communicate with the static service components for client-specific services.
`
`transform
`code for
`performance
`
`collect data
`on program
`behavior
`
`check
`signatures
`execute
`program
`
`204
`
`PALO ALTO NETWORKS Exhibit 1004 Page 3
`
`
`
`centralized services [Spyglass 98].
`A DVM
`introduces a modest amount of new
`functionality into the existing trusted computing base of an
`organization. A DVM client needs to trust that the static
`and dynamic service components it relies on for safety,
`including the proxy and binary rewriter, are implemented
`correctly. In addition, any service authentication scheme
`used in the clients, which may include a digital signature
`checker and a key manager, form part of the trusted
`computing base under our design. However, we believe the
`actual impact of these additions to the TCB to be small.
`Monolithic clients already
`trust all of
`the service
`components that form a traditional VM and often already
`have provisions for cryptographic protocols and digital
`signature checking to support their target applications
`[Gong 99]. Overall, a modest increase in the TCB enables
`DVM clients
`to migrate
`the
`trusted components
`to
`physically secure, professionally managed and administered
`hosts, which is critical to addressing the operational
`problems that have plagued monolithic VMs.
`several
`in
`Our
`service architecture
`is unique
`fundamental ways. First, the centralized services are
`mandatory for all clients in an organization. For example,
`security checks injected into incoming code are inseparable
`from applications at the time of their execution and are thus
`binding throughout the network. Second, there is a single
`logical point of control for all virtual machines within an
`organization. In the case of the security service, policies are
`specified and controlled
`from a
`single
`location;
`consequently, policy changes do not require the cooperation
`of unprivileged users. And third, using binary rewriting as a
`service implementation mechanism preserves compatibility
`with existing monolithic virtual machines. A monolithic
`virtual machine may subject the rewritten code to redundant
`checks or services, but it can take advantage of the added
`functionality without any modifications.
`While a distributed service architecture addresses the
`problems faced by monolithic virtual machines, it may also
`pose new challenges. Centralization can lead to a bottleneck
`in performance or result in a single point of failure within
`the network. These problems can be addressed by
`replicated or recoverable server implementations. The next
`section shows how the separation between static and
`dynamic service components can be used to delegate state-
`requiring functionality to clients. Section 4 shows that this
`implementation strategy does not pose a bottleneck for
`medium sized networks even in the worst possible case and
`can easily be replicated to accommodate large numbers of
`hosts.
`3. Services
`We have implemented the architecture described in the
`previous section to support a network of Java virtual
`machines (JVMs). In
`this section, we describe
`the
`
`implementation of conventional virtual machine services
`under our architecture and show that the distributed
`implementation
`of
`these
`services
`addresses
`the
`shortcomings of monolithic VMs outlined in the first
`section. Our services are derived from the Java VM
`specification, which broadly defines a type-safe, object-
`based execution environment. Typical implementations
`consist of a verifier, which checks object code for type-
`safety, an interpreter and a set of runtime libraries. In some
`implementations, the interpreter is augmented with a just-
`in-time compiler to improve performance. The following
`sections describe the design and implementation of the
`services we have built to supplant those found in traditional
`Java virtual machines.
`All of our services rely on a common proxy
`infrastructure that houses the static service components.
`The proxy transparently intercepts code requests from
`clients, parses
`JVM bytecodes and generates
`the
`instrumented program in the appropriate binary format. An
`internal filtering API allows the logically separate services
`described in this section to be composed on the proxy host.
`Parsing and code generation are performed only once for all
`static services, while structuring the services as independent
`code-transformation filters enables them to be stacked
`according to site-specific requirements [Heidemann &
`Popek 94, O’Malley & Peterson 92]. The proxy uses a
`cache to avoid rewriting code shared between clients and
`generates an audit trail for the remote administration
`console. The code for the dynamic service components
`resides on the central proxy and is distributed to clients on
`demand.
`While the implementation details of our virtual machine
`services differ significantly, there are three common themes
`among all of them:
`• Location: Factoring VM services out of clients and
`locating them on servers improves manageability by
`reducing replicated state, aids integrity by isolating
`services from potentially malicious code and simplifies
`service development and deployment.
`• Service Structure: Partitioning services into static and
`dynamic components can enhance performance by
`amortizing the costly parts of a service across all hosts
`in the local network.
`Implementation Technique: Binary rewriting is used to
`implement services transparently. Binary rewriting
`services can be designed to incur a relatively small
`performance overhead while retaining backward-
`compatibility with existing clients.
`3.1 Verification
`A comprehensive set of safety constraints allows a virtual
`machine to integrate potentially malicious code into a
`privileged base system [Stata & Abadi 98, Freund &
`
`•
`
`205
`
`PALO ALTO NETWORKS Exhibit 1004 Page 4
`
`
`
`Mitchell 98]. Indeed, Java’s appeal for network computing
`stems principally from its strong safety guarantees, which
`are enforced by the Java verifier.
`The task of verifying Java bytecode has been a
`challenge for monolithic virtual machines. First, since the
`Java specification is not formal in its description of the
`safety axioms, there are differences between verifier
`implementations. Verifiers from different vendors differ on
`underspecified issues such as constraints on the uses of
`uninitialized objects, subroutine calls, and cross-validation
`of redundant data in class files. Second, monolithic
`implementations tie the verifier to the rest of the VM,
`thereby prohibiting users from using stronger verifiers
`where necessary. Furthermore, monolithic verifiers make it
`difficult to propagate security patches to all deployed
`clients in a timely manner. As a case in point, 15% of all
`accesses to our web site originate from out-of-date browsers
`with well-known security holes for which many patches
`have been issued. Finally, the memory and processing
`requirements of verification render monolithic VMs
`unsuitable for resource limited clients, such as smart cards
`and embedded hosts [Cohen 97]. Some monolithic virtual
`machines for embedded and resource-limited systems have
`abandoned verification altogether for a restricted extension
`model based on trust [HP 99].
`We address
`these
`shortcomings by decoupling
`verification from the rest of the VM, migrating its
`functionality out of clients into a network verification
`service and centralizing the administration of this service.
`Moving verification out of clients poses some challenges,
`however, because parts of the verification process require
`access to client namespaces and have traditionally required
`close coupling with the client JVM. Specifically, Java
`verification consists of four separate phases. The first three
`operate on a single class file in isolation, respectively
`making sure that the class file is internally consistent, that
`
`the code in the class file respects instruction integrity and
`that the code is type-safe. The fourth phase checks the
`interfaces that a class imports against the exported type
`signatures
`in
`its namespace, making sure
`that
`the
`assumptions that the class makes about other classes hold
`during linking.
`three phases of
`the first
`In our implementation,
`verification are performed statically in a network server,
`while the link-time checks are performed by a small
`dynamic component on the client. This partitioning of
`functionality eliminates unnecessary communication and
`simplifies service implementation. During the processing of
`the first three phases, the verification service collects all of
`the assumptions that a class makes about its environment
`and computes the scope of these assumptions. For example,
`fundamental assumptions, such as inheritance relationships,
`affect the validity of the entire class, whereas a field
`reference affects only the instructions that rely on the
`reference. Having determined these assumptions and their
`scope, the verification service modifies the code to perform
`the corresponding checks at runtime by invoking a simple
`service component (Figure 3). Since most safety axioms
`have been checked by this time, the functionality in the
`dynamic component is limited to a descriptor lookup and
`string comparison. This lazy scheme for deferring link
`phase checks ensures that the classes that make up an
`application are not fetched from a remote, potentially slow,
`server unless they are required for execution.
`The distributed verification service propagates any
`errors to the client by forwarding a replacement class that
`raises a verification exception during its initialization.
`Hence, verification errors are reflected to clients through
`the regular Java exception mechanisms. Since the Java VM
`specification intentionally leaves the time and manner of
`verification undefined except to say that the checks should
`be performed before any affected code is executed, our
`
`class Hello {
` static boolean __mainChecked = false; // Inserted by the verifier
` public static void main() {
` if(__mainChecked == false) { // Begin automatically generated code
` RTVerifier.CheckField(“java.lang.System”,“out”,
`“java.io.OutputStream”);
` RTVerifier.CheckMethod(“java.io.OutputStream”,“println”,
`“(Ljava/lang/String)V”);
`
` __mainChecked = true;
` // End automatically generated code
` }
` System.out.println(“hello world”);
` }
`}
`
`Figure 3. The hello world example after it has been processed by our distributed verification service. The vast majority of safety axioms
`are checked statically. Remaining checks are deferred to execution time, as shown in italics. The first check ensures that the System
`class exports a field named “out” of type OutputStream, and the second check verifies that the class OutputStream
`implements a method, “println,” to print a string. The rewriting occurs at the bytecode level, though the example shows equivalent
`Java source code for clarity.
`
`206
`
`PALO ALTO NETWORKS Exhibit 1004 Page 5
`
`
`
`approach conforms to the specification.
`While this approach to verification does not make the
`central task of verification any easier, it addresses the
`operational problems that have plagued monolithic clients.
`First, it allows a network of VMs to share the same verifier,
`thereby ensuring that code accepted for execution within
`that administrative domain is at least as strong as the
`constraints imposed by the central verifier. Transparent
`binary rewriting ensures that even existing monolithic VMs
`benefit from the assurance provided by the central verifier,
`though they will subject the code to redundant local
`verification. Second, an isolated verification service with
`clear interfaces is easier to check for correctness using
`automatic testing techniques [Sirer & Bershad 99]. Third,
`distributed virtual machine clients can be made smaller and
`cheaper because they do not have to support a verifier.
`Finally, in response to security flaws, only one software
`component needs to be updated per administrative domain,
`in contrast to having to patch every single network client,
`which often requires the assistance of users.
`3.2 Security
`The overall aim of any security service is to enforce a
`specific
`security policy
`across hosts within
`an
`administrative domain. Such a service should meet the
`following three goals to be comprehensive, easy to manage
`and versatile. First,
`it should uniformly enforce an
`organization’s security policy across all nodes. Second, it
`should provide a single point of control for specifying
`organization-wide policies. And, third, it should allow an
`administrator to impose checks in any code deemed
`important for security.
`Our architecture satisfies these goals by factoring most
`functionality that is critical for security out of the individual
`nodes into a centralized network security service. The
`security service forces applications to comply with an
`organization’s security policy by inserting appropriate
`access checks through binary rewriting. The secured
`applications then execute on the individual hosts where a
`small enforcement manager [Grimm & Bershad 99]
`
`Security
`Policy
`
`Security
`Service
`
`Application
`
`Client
`Self-
`Securing
`Application
`Enforcement
`Manager
`Runtime
`Figure 4. The structure of the security service. A master security
`policy determines how applications are rewritten. An enforcement
`manager, residing on the clients, resolves access control checks
`against the central policy.
`
`performs the inserted access checks in accordance with the
`centralized policy (Figure 4).
`Our security model derives from DTOS [Minear 95,
`Olawsky et al. 96, SCC 97], where security identifiers,
`representing protection domains, are associated with
`threads and security-critical objects. Permissions represent
`the right to perform an operation and are associated with
`object methods. An organization-wide policy specification,
`written in a high-level, domain-specific language based on
`XML [Bray et al. 98], specifies an access matrix [Lampson
`71]
`that
`relates security
`identifiers
`to permissions,
`determining who can perform which operation. The policy
`also specifies the mapping between named resources and
`security identifiers, which determines the restrictions on
`named resources such as files, and the mapping between
`security operations and application code, which determines
`where to insert access checks into applications. The security
`service parses
`this policy and accordingly rewrites
`incoming applications, inserting calls to the enforcement
`manager at method and constructor boundaries so that
`resource accesses are preceded by the appropriate access
`checks. During execution of the rewritten application, the
`enforcement manager executes the inserted access checks,
`querying
`the security service based on
`the security
`identifiers and permissions it maintains. As a result, the
`security service performs the bulk of the functionality for
`policy processing and access mediation, reducing client-
`side checking to simple and efficient lookups. The client
`caches the results of security lookups for performance, and
`a cache-invalidation protocol between the security server
`and the enforcement manager enables
`the server
`to
`propagate changes in the access control matrix to clients.
`Policy changes
`that require new code paths
`to be
`instrumented, which we assume to be infrequent, require
`that
`applications be
`restarted using
`the
`remote
`administration service.
`Our design addresses the two fundamental shortcomings
`of the security services in monolithic JVMs [Wallach et al.
`97, Gong 97, Gong et al. 97]. First, in these systems the
`security policy is specified and enforced separately on each
`individual node. Consequently, a site administrator must
`maintain the security state for each node individually, which
`takes an effort proportional to the number of nodes.
`Second, the capability and stack-introspection schemes
`used in monolithic JVMs tie the security policy to the
`implementation of the system and therefore necessitate
`assistance from the original system developers to work.
`Essentially, the developers must anticipate the security
`requirements of the system and embed the requisite security
`checks (or hooks for such checks) at the appropriate places
`in the system libraries. In the case of Java, the developers of
`popular JVMs have anticipated and provided security
`checks on file, network and system property accesses.
`However, audio devices and window creation events have
`
`207
`
`PALO ALTO NETWORKS Exhibit 1004 Page 6
`
`
`
`not been adequately protected, giving rise to denial-of-
`service attacks [McGraw & Felten 99]. In general, access
`control mechanisms that depend on a-priori anticipation and
`manual instrumentation are likely to be less flexible and
`less secure than systems that provide a clear separation
`between security policy and implementation, and allow the
`policy to be specified after deployment by users [Levin et
`al. 75, Erlingsson & Schneider 99].
`3.3 Remote monitoring
`System administration is particularly challenging for large
`networks of heterogeneous systems. Tasks such as resource
`accounting and usage pattern analysis, while easy to
`perform on time-shared systems of the past, are increasingly
`difficult to carry out in today’s distributed systems. Our
`remote monitoring service allows administrators to track
`network-wide resource usage. Patterned after the security
`service,
`the
`remote monitoring
`service
`transforms
`applications to invoke auditing services at the appropriate
`places, such as on entry to and exit from method and
`constructor calls. As each application comes up, it contacts
`the remote monitoring console and a handshake protocol
`establishes the credentials of the user and assigns an
`identifier to the session. This connection is then used to
`send information to a central administration host that
`monitors client hardware configurations, users, JVM
`instances, code versions and noteworthy client events. Since
`the audit logs are stored on an external host that is not
`exposed to untrusted applications, a security breach may
`stop the creation of new audit events but cannot tamper
`with existing audit logs.
`In addition to method-level remote monitoring services,
`we provide an instruction-level profiling and tracing service
`for monitoring application performance. The profiler
`instruments code to generate a dynamic call-graph [Graham
`et al. 82] from applications running across the network, as
`well as statistics on client code usage. We have used these
`services extensively both to debug our system and for
`optimizations. In particular, we have used the tracing
`service to obtain traces of synchronization behavior for
`Java applications and utilized this data in designing a
`transparent optimization service [Aldrich et al. 99].
`3.4 Compilation
`Compilation i