Folks often ask me how to approach a suspected performance issue. My personal strategy is informed by the fact that I work on concurrency issues. (When you have a hammer everything looks like a nail, but I'll try to keep this general). A good starting point is to ask yourself if the observed performance matches your expectations. Expectations might be derived from known system performance limits, prototypes, and other software or environments that are comparable to your particular system-under-test. Some simple comparisons and microbenchmarks can be useful at this stage. It's also useful to write some very simple programs to validate some of the reported or expected system limits. Can that disk controller really tolerate and sustain 500 reads per second? To reduce the number of confounding factors it's better to try to answer that question with a very simple targeted program. And finally, nothing beats having familiarity with the technologies that underlying your particular layer.
On the topic of confounding factors, as our technology stacks become deeper and less transparent, we often find our own technology working against us in some unexpected way to choke performance rather than simply running into some fundamental system limit. A good example is the warm-up time needed by just-in-time compilers in Java Virtual Machines. I won't delve too far into that particular hole except to say that it's rare to find good benchmarks and methodology for java code. Another example is power management on x86. Power management is great, but it can take a while for the CPUs to throttle up from low(er) frequencies to full throttle. And while I love "turbo" mode, it makes benchmarking applications with multiple threads a chore as you have to remember to turn it off and then back on otherwise short single-threaded runs may look abnormally fast compared to runs with higher thread counts. In general for performance characterization I disable turbo mode and fix the power governor at "performance" state. Another source of complexity is the scheduler, which I've discussed in prior blog entries.
Lets say I have a running application and I want to better understand its behavior and performance. We'll presume it's warmed up, is under load, and is an execution mode representative of what we think the norm would be. It should be in steady-state, if a steady-state mode even exists. On Solaris the very first thing I'll do is take a set of "pstack" samples. Pstack briefly stops the process and walks each of the stacks, reporting symbolic information (if available) for each frame. For Java, pstack has been augmented to understand java frames, and even report inlining. A few pstack samples can provide powerful insight into what's actually going on inside the program. You'll be able to see calling patterns, which threads are blocked on what system calls or synchronization constructs, memory allocation, etc. If your code is CPU-bound then you'll get a good sense where the cycles are being spent. (I should caution that normal C/C++ inlining can diffuse an otherwise "hot" method into other methods. This is a rare instance where pstack sampling might not immediately point to the key problem). At this point you'll need to reconcile what you're seeing with pstack and your mental model of what you think the program should be doing. They're often rather different. And generally if there's a key performance issue, you'll spot it with a moderate number of samples.
I'll also use OS-level observability tools to lock for the existence of bottlenecks where threads contend for locks; other situations where threads are blocked; and the distribution of threads over the system. On Solaris some good tools are mpstat and too a lesser degree, vmstat. Try running "mpstat -a 5" in one window while the application program runs concurrently. One key measure is the voluntary context switch rate "vctx" or "csw" which reflects threads descheduling themselves. It's also good to look at the user; system; and idle CPU percentages. This can give a broad but useful understanding if your threads are mostly parked or mostly running. For instance if your program makes heavy use of malloc/free, then it might be the case you're contending on the central malloc lock in the default allocator. In that case you'd see malloc calling lock in the stack traces, observe a high csw/vctx rate as threads block for the malloc lock, and your "usr" time would be less than expected.
Solaris dtrace is a wonderful and invaluable performance tool as well, but in a sense you have to frame and articulate a meaningful and specific question to get a useful answer, so I tend not to use it for first-order screening of problems. It's also most effective for OS and software-level performance issues as opposed to HW-level issues. For that reason I recommend mpstat & pstack as my the 1st step in performance triage. If some other OS-level issue is evident then it's good to switch to dtrace to drill more deeply into the problem.
Only after I've ruled out OS-level issues do I switch to using hardware performance counters to look for architectural impediments.