Scaling dcache with RCU
Although this change in dcache was relatively small, it had far-reaching consequences in the kernel, because a well-defined API for filesystems to interact with dcache was not in place. This resulted in a large number of bugs in the Linux 2.5 kernel due to filesystems hackers attempting to manipulate dcache directly in the traditional style. Given that a somewhat more formal API now exists, we hope future changes will be less traumatic.
Figure 9 shows the performance of a multiuser benchmark running on a Linux 2.5.59 kernel patched to use RCU in the directory-entry cache compared to the performance of an unpatched kernel. These benchmarks were run on a 16-CPU NUMA-Q system using 700MHz PIII Intel Xeons with 1MB L2 cache and 16GB of memory.
Applying the dcache_rcu patch to a Linux 2.4.17 kernel increased SPECweb99 (without SSL) throughput from 2,258 to 2,530 on an 8-CPU PIII Xeon server, a 12% improvement. Applying the same patch to a Linux 2.5.40-mm2 kernel reduced the system time consumed by a Linux kernel build from 47.548 CPU seconds to 42.498 CPU seconds, more than a 10% reduction. A similar test run on a uniprocessor 700MHz PIII Xeon system running the Linux 2.5.42 kernel showed no change. In summary, dcache RCU not only increases scaling for high-end machines, it also maintains good performance on low-end machines.
Although the 2.6 dcache system is much more scalable than the 2.4 version was, a number of issues still need to be investigated:
Updates still are gated by dcache_lock, which means that update-intensive workloads do not scale well.
The global hash table defeats cache locality and makes update code more complex than necessary. Of course, any alternative must preserve its benefits, including high-performance handling of large directories.
The 2.6 dcache code acquires each dentry's d_lock spinlock, resulting in cache-line bouncing and atomic operations, particularly on the root directory and on working directories. Much thought is needed to arrive at a simple solution, as moving permissions into the dentry turns out to be quite complex.
The code that resolves races between __d_lookup() and d_move() is overly complex.
We eagerly anticipate participating in the 2.7 effort to resolve these issues, hopefully resulting in the situation shown in Figure 10.
This work represents the view of the author and does not necessarily represent the view of IBM.
SPEC and the benchmark name SPECweb are registered trademarks of the Standard Performance Evaluation Corporation. The benchmarking was done for research purposes only and may not be compared to published results on the SPECWeb site, due to the following deviations from the rules:
It was run on hardware that does not meet the SPEC availability-to-the-public criteria. The machine was an engineering sample.
access_log was not kept for full accounting. It was being written but deleted every 200 seconds.
For the latest SPECweb99 benchmark results, visit www.spec.org.
Paul E. McKenney is a distinguished engineer at IBM and has worked on SMP and NUMA algorithms for longer than he cares to admit. Prior to that, he worked on packet-radio and Internet protocols (but long before the Internet became popular). His hobbies include running and the usual house-wife-and-kids habit.
Dipankar Sarma currently is working on a number of Linux kernel projects, including CPU hot-plug, RCU and VFS enhancements. Prior to his Linux days, he worked on a number of areas including ABI, OS bringup, I/O drivers and multipath I/O.
Maneesh Soni has been working with IBM's Linux Technology Center as a member of Linux Scalability Effort Project. He has experience in the system software arena, particularly with operating-system kernels and filesystems.
|Nativ Disc||Sep 23, 2016|
|Android Browser Security--What You Haven't Been Told||Sep 22, 2016|
|The Many Paths to a Solution||Sep 21, 2016|
|Synopsys' Coverity||Sep 20, 2016|
|Naztech's Roadstar 5 Car Charger||Sep 16, 2016|
|RPi-Powered pi-topCEED Makes the Case as a Low-Cost Modular Learning Desktop||Sep 15, 2016|
- Android Browser Security--What You Haven't Been Told
- Nativ Disc
- The Many Paths to a Solution
- Recovery of RAID and LVM2 Volumes
- Download "Linux Management with Red Hat Satellite: Measuring Business Impact and ROI"
- Naztech's Roadstar 5 Car Charger
- Synopsys' Coverity
- Securing the Programmer
- RPi-Powered pi-topCEED Makes the Case as a Low-Cost Modular Learning Desktop
- Glass Padding
With all the industry talk about the benefits of Linux on Power and all the performance advantages offered by its open architecture, you may be considering a move in that direction. If you are thinking about analytics, big data and cloud computing, you would be right to evaluate Power. The idea of using commodity x86 hardware and replacing it every three years is an outdated cost model. It doesn’t consider the total cost of ownership, and it doesn’t consider the advantage of real processing power, high-availability and multithreading like a demon.
This ebook takes a look at some of the practical applications of the Linux on Power platform and ways you might bring all the performance power of this open architecture to bear for your organization. There are no smoke and mirrors here—just hard, cold, empirical evidence provided by independent sources. I also consider some innovative ways Linux on Power will be used in the future.Get the Guide