Tuning Toolbox

Over the past several years, the Linux Kernel Performance Project [1] has tracked the performance of Linux and tuned it for throughput and power efficiency on Intel platforms. This experience has given us some insights into the best tools and techniques for tuning Linux systems. In this article, we describe some of our favorite Linux performance utilities and provide a real-world example that shows how the Kernel Performance Project uses these tools to hunt down and solve a real Linux performance issue.

Finding Bottlenecks

The first task in performance tuning is to identify any bottlenecks that might be slowing down system performance.

The most common bottlenecks occur in I/O, memory management, or the scheduler. Linux offers a suite of tools for examining system use and searching out bottlenecks. Some tools reveal the general health of the system, and other tools offer information about specific system components.

The vmstat utility offers a useful summary of overall system performance. Listing 1 shows vmstat data collected every two seconds for a CPU-intensive, multi-threaded Java workload. The first two columns (r, b) describe how many processes in the systems can be run if a CPU is available and how many are blocked. The presence of both blocked processes and idle time in the system is usually a sign of trouble.

The next four columns under memory show how much memory space is used. Frequently swapping memory in and out of the disk swap space slows the system. The cache column gives the amount of memory used as a page cache. A bigger cache means more files cached in memory. The two columns under io, bi, and bo, indicate the number of blocks received and sent to block devices, respectively, which gives an idea of the level of disk activity. The two columns under system, in, and cs, reveal the number of interrupts and context switches.

If the interrupt rate is too high, you can use an interrupt utility, like sar, to help uncover the cause. The command sar -I XALL 10 1000 will break down the source of the interrupts every 10 seconds for 1000 seconds. A high number of context switches relative to the number of processes is undesirable because of flushing of cached data.

The next four columns in Listing 1, us, sy, id, and wa, indicate the percentage of time the CPU(s) has spent in userspace applications, in the kernel, being idle, or waiting for I/O, respectively. This output shows whether the CPUs are doing useful work or whether they are just idling or being blocked. A high percentage of time spent in the OS could indicate a non-optimal system call. Idle time for a fully loaded system could point to lock contentions.

Disk Performance

Hdparm is a good tool for determining whether the disks are healthy and configured:

# hdparm -tT /dev/sda
/dev/sda:
Timing buffered disk reads: 184 MB in 3.02 seconds = 60.88 MB/sec
Timing cached reads: 11724 MB in 2.00 seconds = 5870.80 MB/sec

The preceding command displays the speed of reading through the buffer cache to the disk, with and without any prior caching of data. The uncached speed should be somewhat close to the raw speed of the disk. If this value is too low, you should check in your BIOS to see whether the disk mode is configured properly. Also, you could check the hard disk parameter setting for an IDE disk

# hdparm -I /dev/hda

or for a SCSI disk:

# sdparm /dev/sda

To study the health of a run-time workload's I/O, use iostat. For example, Listing 2 shows how to use iostat for dumping a workload. If %iowait is high, CPUs are idle and waiting for outstanding disk I/O requests. In that case, try modifying the workloads to use asynchronous I/O or dedicate a thread to file I/O so workload execution doesn't stop.

The other parameter to check is the number of queued I/O requests: avgqu-sz. This value should be less than 1 or disk I/O will significantly slow things down. The %util parameter also indicates the percentage of time the disk has requests and is a good indication of how busy the disk is.

CPU Cycles

One important way to identify a performance problem is to determine how the system is spending its CPU cycles. The oprofile utility can help you study the CPU to this end. Oprofile usually is enabled by default. If you compile your own kernel, then you need to make sure that the kernel configs CONFIG_OPROFILE=y and CONFIG_HAVE_OPROFILE=y are turned on.

The easiest way to invoke oprofile is with the oprofile GUI that wraps the command-line options. To do so, use oprofile 0.9.3 or later for an Intel Core 2 processor and install the oprofile-gui package. Now invoke

#oprof_start

to bring up the Start profiler screen with Setup and Configuration tabs (Figure 1). First, select the Configuration tab. If you want to profile the kernel, enter the location of the kernel image file (that is, the uncompressed vmlinux file if you compile the kernel from source). Now return to the Setup tab.

Figure 1: Profiling the kernel with oprofile.

In the Events table, select the CPU_CLK_UNHALTED event and the unit mask Unhalted core cycles. Note: Normally, you do not need to sample the system any more often than the setting listed under in the Count field.

A lower count means that fewer events will need to happen before a sample is taken, thus increasing the sampling frequency. Now run the application you want to profile, and start oprofile by clicking on the Start button. When the application has stopped running, click the Stop button.

To view the profile data, invoke:

#opreport -l

The output for this command is shown in Listing 3.

Listing 3 shows the percentage of CPU time spent in each application or kernel, and it also shows the functions that are being executed. This report reveals the code the system is spending the most time in, which should improve performance if you can use this data as a basis for optimization.

If you have collected call graph information, type the command

#opreport -c

to obtain the output shown in Listing 4.Listing 4 shows that this workload has some very heavy memory allocation activity associated with getting free memory pages and clearing them.

Too Many Cache Misses?

The performance of the system is highly dependent on the effectiveness of the cache. Any cache miss will degrade performance and lead to a CPU stall.

Sometimes a cache miss is caused by frequently used fields located in data structures that span across the cache line. Oprofile can diagnose this kind of problem.

Again, using the Intel Core 2 processor as an example, choose the event LLC_MISSES to profile all the L2 cache requests that miss the L2 cache. For the exact event to use, you should invoke opcontrol --list-events to read about the details of each event type available for your CPU.

Listing 5 shows how to call up a cache miss profile.

Oprofile is a very versatile tool. By carefully choosing which events to monitor, you can zero in on the CPU operation that is causing the problem.

Locking Problems

A high context switching rate, relative to the number of running processes, is undesirable and could indicate a lock contention problem. To determine the most contended locks, enable the lock statistics in the kernel, which will give you insight into what is causing the lock contention. To do so, use the lock_stat feature in 2.6.23 or later kernels. First, you'll need to recompile the kernel with the CONFIG_LOCK_STAT=y option. Then, before running the workloads, clear the statistics with:

#echo 0 > /proc/lock_stat

After running the workload, review the lock statistics with the following command:

#cat /proc/lock_stat

The output of the preceding command is a list of locks in the kernel sorted by the number of contentions. For each lock, you will see the number of contentions, as well as the shortest, maximum, and cumulative wait time for a contention. In addition, you will see the number of acquisitions, as well as the minimum, maximum, and cumulative hold times for a lock. The top call sites of the lock are also given to let you locate quickly where in the kernel the lock occurs.

It is worth noting that the lock statistics infrastructure incurs overhead. Once you have finished hunting for locks, you should disable this feature to maximize performance.

Excessive Latency

Program throughput that is inconsistent and sputters, applications that seem to go to sleep before coming alive, and a lot of processes under the blocked column in vmstat are often signs of latency in the system. LatencyTOP is a new tool that helps diagnose latency issues.

Starting with the 2.6.25 kernel, you can compile LatencyTOP support into the kernel by enabling the CONFIG_HAVE_LATENCYTOP_SUPPORT=y and CONFIG_LATENCYTOP=y options in the kernel configuration. After booting up the kernel with LatencyTOP capability, you can trace latency in the workload with a userspace latency tracing tool from the LatencyTOP website [2]. To start, compile the tool, do a make install of the LatencyTOP program, and run the following as root:

#./latencytop

The LatencyTOP program's top screen (Figure 2) provides a periodic dump of the top causes that lead to processes being blocked, sorted by the maximum blocked time for each cause. Also, you'll find information on the percentage of time a particular cause contributed to the total blocked time. The bottom screen provides similar information on a per-process basis.

Figure 2: Studying system latency with LatencyTOP.

An Example

Linux provides quick allocation and deallocation of frequently used objects in caches called "slabs." To provide better performance, Christopher Lameter introduced a new slabs manager called Slub.

However, we found that the scheduler performance benchmark known as hackbench reveals a big difference in run time with kernel 2.6.24/2.6.25-rc, between a system with 16 CPU cores and a system with eight CPU cores. Hackbench is expected to be faster on the 16-core system than on the 8-core system, but the testing result shows the first machine requires three times more run time than the second machine, which indicates a possible performance issue.

The vmstat utility provides the output shown in Listing 6.

Notice the high context switch (cs) count and large number of running processes. In this case, hackbench simulates many chat rooms with a large number of users passing messages back and forth in each room. The lack of idle time in the system indicates that the CPU is very busy.

The next step is to use oprofile to find out where the CPU is spending its time. The oprofile data in Listing 7 shows that about 88% of the CPU time is spent in allocating slabs, adding to partially filled slabs, and freeing slabs. It shows that the benchmark generates lots of messages that are allocated and passed between processes and memory management, and that is where the program is spending the most time.

This result indicates the need to take a closer look at what is going on with the slabs. A utility called slabinfo provides a report on slab activity. (The source code for the slabinfo utility is with the kernel source under Documents/vm/slabinfo.c.) To obtain information about the most actively used objects, invoke the slabinfo utility (see Listing 8).

The block objects, size 192 and 512, are actively used by hackbench messages: One is for the socket buffer header and one is for the message body.

Basically, the SLUB implementation keeps a per-cpu cache for each slab type. When the kernel allocates an object, it checks the per-cpu cache first without locking. Such allocation is very fast and is called a fast path. If the per-cpu cache hasn't freed objects, the kernel allocates from shared pages with a lock, which is slow. A slow path means more lock contentions. The free procedure also has a fast path and a slow path. Because free uses a distributed lock (page lock) and the allocation process uses more exclusive locks, allocation by fast path is more important.

For these two objects, we noted that the free operation is quite slow; however, allocation is not fast, either. For example, for objects of size 512, only 68% of allocation is by fast path, and 7% of free is by fast path.

To reduce the slow path allocation, we could ask for a bigger sized slab to increase the per-cpu object cache. To increase the default max_order of 1 and min_objects of 32, we add slub_max_order=3 slub_min_objects=32 to the kernel boot command line. This increases the number of objects that must fit into one slab for an allocation to be successful, which will reduce the chance that the kernel allocates objects by slow path.

This step improved the throughput significantly, requiring just one tenth the time needed in the previous test. By extensive testing with different slub_min_objects settings, we found the correlation between slub_min_objects and the CPU number.

Mostly, we get the best result with slum_min_objects=cpu_number*2. If slum_min_objects is equal to a bigger value, the result doesn't provide much improvement.

At this point, we went back to the 8-core machine and did extensive testing to confirm our findings. After we discussed the problem with the SLUB maintainers, a patch that scales slub_min_objects, as a function of the number of CPU cores, was merged into the Linux kernel.

Conclusions

In this article, we provided a quick tour of some useful tools for diagnosing common performance issues. Of course, this brief introduction is not intended as a comprehensive description of the performance tuning craft, but it should provide you with a good starting point for discovering and fixing performance bottlenecks on your Linux systems.

# cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_governors
ondemand userspace performance

# cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

# echo ondemand > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor