Monitoring system and network performance metrics
Shrink-wrapped systems management solutions just don't cut it. We show you how we customized a fault and performance management system for one company to improve availability and response time for customers and internal users
Our ability to construct complex applications and networks has far outstripped our ability to manage these environments. The application and network vendor communities try to convince us that their products are readily adaptable to any environment and are fully manageable with embedded utilities or standalone monitoring software. It is naive to believe shrink-wrapped solutions can come close to providing the level of management necessary to satisfy end user requirements.
Here's how one company evolved their enterprise management system to meet the challenges of distributed computing. (2,900 words)
orporate and information technology management have followed the lead of the pointy-haired boss in the Dilbert cartoon in oversimplifying the problem of designing, implementing, and especially, supporting enterprise client/server applications. The typical IS department is now forced down one of two paths to deal with the problem -- don't make waves, chuckle at relevant Dilbert strips, and let the support fantasy continue or document the real support requirements and be prepared to defend your position. The don't-make-waves scenario is akin to an ostrich on a busy freeway, where the operational support problems will run you over even if you bury your head. This article provides no insight for those who choose the ostrich mode of operational support.
If, however, you are among the few brave persons willing to stand your ground and fight for the tools you need to really support your environment, then you may find this "real world perspective" useful. It has been our experience that corporate managers will respond positively to system/network system requests provided alternatives are presented and a brief ROI analysis is submitted. Alternatives can be as straightforward as conducting a lost business or service analysis (i.e., if the application, network, or server goes down for "X" amount of time, what is the real world "$$$$$" impact on the business?). When conducting this type of analysis, remember to consider the following cost points:
In the information age, time is even more important than money.
Operating environment and requirements
Enterprise management systems need to meet a number of management criteria. The OSI provides a definition that quantifies management functions into discrete buckets: fault, performance, configuration, security, and accounting. Although a full discussion of each of these management functions would require far more words than allocated to this article, a case study that focuses on some fault and performance-management issues may be helpful in your justification process.
A magazine publishing firm (Enterprise A) decided to make a bold move by offering online subscriptions of their published periodicals to Internet subscribers. This offering included reformatting their periodicals into .PDF file types and building a large Netscape server that could access periodicals stored on Oracle databases while providing reliable access to Internet subscribers. The primary online publication environment was comprised of four Sun database servers, two Sun development servers, and a large 6-way Sun Netscape server connected through a routed network comprised of several Ethernet and fast Ethernet segments. The corporate operational environment shared the network infrastructure and was comprised of 800 Novell and Unix users with 18 Novell servers and 10 Solaris-based application servers distributed across four sites.
The task at hand was to implement a management system that could provide fault and performance management for the online publication systems and corporate end-user environment. The emphasis of the system was to be able to provide high levels of service to online subscribers and internal users.
After meeting with the end users and the online support team, it became readily apparent that the primary business requirements for the management system consisted of two metrics: availability and response time. Armed with this information, our company, the Netplex Group, designed and installed a platform-based management system that provided the primary fault and performance management functions needed to meet these main service objectives. The management system was comprised of the following components:
Network Node Manager -- the platform
The 4.1 version of the Network Node Manager (NNM) platform supports over 300 third-party management applications (such as CiscoWorks, 3Com's Transcend, etc.), which allows the platform to become the foundation for most element management requirements. The 4.1 NNM platform is based on a fully distributed model that supports up to 15 domain managers and 15 management consoles. Events filters can be created to allow enterprise-based information to be forwarded to the master console. The distributed architecture will prove useful with future implementations since each of the main sites has enterprise and local support functions.
PerfView/MeasureWare -- The system/application performance monitor suite
The PerfView/MeasureWare product suite provides correlated threshold based performance management for Unix, Novell 3.12 and 4.1, and NT 3.51 and 4.0 servers. The product suite is comprised of several components: the server agent, which resides on the file servers; the MeasureWare agent, which resides on a Unix server and functions as an intermediate collection facility, a Unix station agent, and a correlated data source collector; and the application suite which must run on an HP-UX or NT server and includes the PerfView analyzer, monitor, and trending modules. The applications can provide on-screen and formatted reports of both real-time and historical data. Historical data can be run through statistical trending algorithms to provide capacity forecasting trends. The PerfView/MeasureWare product suite is immensely scalable in both architecture and data integration. The middle-level function of the MeasureWare agent allows mid-level data collection facilities to be separated or collocated with the management console. The standards-based Data Source Integration (DSI) allows any SNMP- or ASCII-based data to be integrated and correlated to other MeasureWare parameters. The customized NetWare DSI template was used to integrate the Novell SNMP data into the MeasureWare agent.
NetMetrix -- the network performance monitor suite
HP NetMetrix is an application that provides for remote monitoring of Ethernet, Token ring, FDDI, and Fast Ethernet network segments. RMON/RMON2 metrics that can be monitored include utilization, various topology-based errors, packet size, and traffic generated and received from specific network devices. NetMetrix monitors the health of network segments and generates threshold-based events to a management console. NetMetrix also has a protocol analyzer tool that can be started via an application icon or triggered by a network event threshold. Using Internet standard Echo MIB, NetMetrix can perform response time monitoring of data packets for different protocols. The Echo MIB has been integrated into the standard HP RMON probes and is available as a server Internetwork response time agent (IRA). The data collected by NetMetrix can be displayed on-screen, integrated into printed reports or accessed via the Web.
Onion Peel NDC -- the enhanced SNMP collection utility
The Onion Peel NDC software augments the base functionality of the NNM platform. NDC is a fully integrated NNM software tool that is comprised of software which facilitates the class-based collection of MIB data. The class-based collection that supports wild cards and SQL-like filters allows for general MIB object classes to be specified for SNMP collection as opposed to manufacturer specific MIBs. The NDC tool has a flexible output formatting function which can also be used to import collected data into high-end modeling tools like SAS, report generation engines, or spreadsheets.
Onion Peel ROVE -- the ROVE SNMP-enhanced, multiuser, and remote access utility
The Onion Peel ROVE (remote OpenView environment) product provides enhanced remote access to the base NNM platform. ROVE is a fully integrated NNM software tool that facilitates remote access into the management station without incurring the process and communication burden of remote X Windows sessions. ROVE Motif and VT-100 clients provide a mechanism that allows support staff to remotely access the management server from remote Motif and terminal clients respectively. The Motif clients are configured to share a common X-terminal session which limits the performance impact on the management server. The VT terminal client provides a color-coded tabular interface to critical OpenView data (object states, traplog, etc.) for LAN and dial-based terminal connections. The VT-100 emulation is especially important to on-call IS personnel that live a distance from the data center. A page in the middle of the night can be assessed and many times resolved remotely.
The HP 9000/871 model D370 Server was installed and configured with a mirrored boot partition to enhance the availability of the enterprise management server. A color PostScript printer was directly connected to HP management server ensure that cron-based reports could be printed. With the server hardware fully operational the NNM platform software and associated patches were loaded. Remember, there are almost always patches out there that will make your life a little easier. As previously specified, the platform software needed to be installed first due to dependencies of the subordinate element applications. We next loaded the correct vendors' MIB versions, the most important being the Cisco router MIB. With the MIBs loaded, seed router addresses were quantified and the autodiscovery process was run. We were now ready to load our applications.
The NetMetrix application modules were loaded and integrated into the NNM platform. Nine critical segments were identified and instrumented with RMON2 probes (intelligent LAN monitoring agents). Once implemented, a two-week baseline was run for each segment to determine current operating levels. Note that it was important to monitor whether operational levels were viewed as acceptable throughout the baseline period. Segment-specific baselines were taken for protocol-specific (ICMP and Novell 802.3) response time, utilization, broadcasts, and errors. The combination of utilization, broadcasts, response time, and errors helped quantify network availability, while the protocol-specific response time clearly delineated network/system responsiveness.
It should noted that during the initial baseline, performance was viewed as unacceptable, but the data was used to justify the addition of a FastEthernet backbone segment. With network operations brought to an acceptable level, a baseline was run a second time and threshold-based alarms were set for each of the metrics on each of the segments. Specific cron-based reports and Web-based display formats were implemented for the monitored LAN segments. The combination of the threshold-based notification, comprehensive diagnostic tools, cron-based reports, Web-based interface, and remote packet capture capability allowed the IS staff to proactively support the network infrastructure.
The next focus was on the integration of the Onion Peel NDC product and enhancement of the NNM platform to provide better management of the Cisco routers. The NDC utility was used to provide a historical SNMP collection of 30 of the 5,000+ Cisco MIB variables. The output of the collection was stored in tabular flat file, which was then exported to an Excel spreadsheet on a daily basis. A custom application was created via the NNM application builder to view the same 30 MIB variables on a real-time basis. The NNM tool menu was also expanded to include a router interface reset function, which allowed the support staff to remotely reset any router interface in the network. The combination of the native Cisco traps, customized tool bar, NNM diagnostic tools (Traceroute, ping, remote ping, ARP cache display, etc.), historical date capture, and real-time views allowed the IS team to proactively manage and thereby improve availability of the Cisco routers.
Now that the network was reasonably solid, we turned our efforts toward management of the systems. The PerfView application was loaded and integrated into the NNM platform. Armed with a system/application performance monitoring application, we focused on instrumenting the servers that were critical to business operations. Avanti SNMP agents were installed on the 18 Novell file servers. The 18 remote data sources were integrated into the MeasureWare agent running on the HP server via the DSI (data source integration) function.
The alarm definition file and NOS connectivity module installed on the HP server were configured to forward threshold-based NetWare alarms to the OV_trap log on the NNM platform. The PerfView applications graphs template were created and displayed critical server functions (CPU_util, Memory_util, Disk_rate, Disk_util, Network_rate, etc.) The combination of the baselined performance metrics and the NNM discovery process provided the quantifiable availability for the Novell file servers. The response time requirement was addressed by including the Novell file servers in the NetMetrix (ICMP and Novell_802.3) response time target lists.
The 17 Sun servers and the HP management server were instrumented with
dedicated MeasureWare agents. These agents can track over 300
metrics sampled from the Kmem portion of the kernel -- the same
portion of the kernel sampled by
The sampling interval is
every five to 10 milliseconds, the process write interval is 1/min,
and the global and application write interval is 5/min.
The Unix-based MeasureWare agents were tuned by shutting off the short-lived process sampling in the parameter definition file because these processes typically do not have significant value from a monitoring perspective and can significantly increase processor utilization. We noticed high CPU utilization (13 to 15 percent) on the Sun servers even after shutting of the data logging of short-lived processes. Upon further investigation, we opted to load the new TOPs process patch, which reduced the MeasureWare process utilization to a cozy (two to three percent). We also grouped the Oracle and Netscape processes into application groups and monitored the application groups as separate entities so as to better quantify the metrics directly associated with these critical applications.
Although this level of monitoring proved to be adequate for the Oracle servers, the Netscape server would sporadically have one or more processes that would fall into a zombie state. The MeasureWare agent could detect missing processes but could not, in the default configuration, detect zombie processes. Given the business-critical function of the six-way Netscape server, we elected to make use of the DSI functionality once again. A client application was tied to a Unix cron process to periodically log into the Netscape server. If the login failed, an exit code was generated and forwarded into the DSI logging process. An automation and custom alarm script was developed for the MeasureWare agent on the Netscape server that would log and alarm on failed login attempts and automatically restart the Netscape server processes if two login attempts failed.
Upon completion of the MeasureWare customization for the Netscape server, PerfView application graph templates were created and used to display critical server and application functions (CPU_util, Memory_util, Memory_page_rate, Disk_rate, Disk_util, Network_rate, etc.) The combination of the baselined performance metrics, application group definitions, customized Netscape monitoring scripts, and the NNM discovery process provided the quantifiable availability for all of the Sun application and database servers. The response time requirement was addressed by including the Sun file servers in the NetMetrix ICMP response time target lists.
Once all the critical monitoring applications were operational we needed to integrate the Onion Peel ROVE software and customize the mailer interface on the NNM platform to improve accessibility to and from the management station. Given the critical nature of an online publication application, many of the IS support people are on call. Before the management system was installed, their only course of action when they received a call from a help desk operator was to drive into the site and repair the system. Needless to say, there were more then a few false alarms. The integration of ROVE into the NNM platform allowed the IS support team to remotely dial into the management server and assess the severity of the problem. The integration of the NNM into the sendmail daemon allowed an e-mail message to be generated in the paging system for critical Netscape, Oracle, router, or segment availability problems. This level of integration was fine tuned to eliminate extraneous pages and proved to be a valuable tool for improving availability.
As you would expect, all the fault and performance management tools that were integrated into the NNM platform needed to be coupled with processes in order to gain the maximum support benefit. This support process topic was introduced in the April 1997 SunWorld Real World Perspectives column, "Cut workplace frustration -- take these steps to build an efficient help desk."
Based on the requirements of this online firm, future selection and integration of trouble ticketing and event correlation systems is being considered. The benefit of selecting a management platform that supports a wide variety of management applications will become readily apparent as there are a number of trouble ticketing and event correlation vendors that integrate into the OpenView NNM platform. The integration of these tools will further enhance the support of existing and future business-critical systems.
Network and systems administrators often are made to feel like living scar tissue constantly in fire-fighting mode. By developing applications performance toolsets on top of enterprise networked systems management platforms, network and systems managers can become proactive members of the business team that provide return on investment by deploying mission-critical systems more efficiently and offering superior service levels to end users.
About the author
Frank Henderson is chief technology officer at The Netplex Group. His expertise is in designing and installing networks and reengineering help desks, and in ORB, distributed databases, and network management. Dave Koehler is director of network strategies at The Netplex Group. He has 13 years experience in distributed network design. Koehler has written for several journals and has delivered papers at ComNet, IEEE, and SunWorld Expo. Reach Frank Henderson at firstname.lastname@example.org and Dave Koehler at email@example.com.
If you have technical problems with this magazine, contact firstname.lastname@example.org