Files
FAT-Allocator/docs/evaluation/evaluation.org
2025-02-10 13:37:59 +00:00

314 lines
19 KiB
Org Mode

* Evaluation
#+BEGIN_COMMENT
We tested the FAT Pointer based range addresses
against Jemalloc the default memory allocator
for CHERIBSD. We evaluate the general improvement
in performance such as wall clock runtime by
reducing the TLB misses by designing a CHERI
based huge page aware allocator. There are
2 classes of benchmarks proposed for
evaluating the proposed allocator against the
system allocator. The 2 classes are mirco and macro
benchmarks. The micro benchmark refers to the specific
set of smaller C programs designed to test certain
specific allocator patterns. The macro benchmark
refers to larger set of C programs to evaluate real
world programs. The sections listed below diveldge
into the following:
- Expirement setup which talks about the software stack
used for evaluating the benchmark.
- Expanding on the classes of C programs executed
and describing the characterics of each of the
c programs.
- Listing the results and describing the behavoir of the
evaluated results.
- Describing the usability of the proposed allocator based
on the evaluated results and limitations identified.
#+END_COMMENT
We conducted tests of the FAT Pointer-based range addresses against Jemalloc\cite{jemalloc},
the default memory allocator for CHERIBSD\cite{cheribsd}, to assess the performance improvements
enabled by a CHERI-based huge page-aware allocator. Specifically, we evaluated
the reduction in TLB misses and its impact on overall
performance metrics, such as wall clock runtime.
To comprehensively analyze the proposed allocator, we categorized benchmarks into
two classes which are micro and macro benchmarks. Micro benchmarks comprise smaller
C programs designed to target specific allocator patterns, enabling us to evaluate
detailed aspects of the allocator's behavior. Macro benchmarks, on the other hand,
encompass larger, real-world C programs, allowing us to assess the allocator's
performance in more practical, real-world scenarios.
The experiment setup section details the software stack used for evaluation. It includes
the specific configurations, compiler options, and system environment tailored
to benchmark the proposed allocator. This ensures consistency and repeatability
in our results, providing a solid foundation for meaningful comparisons.
We further elaborated on the two classes of benchmarks executed. Micro benchmarks
focused on particular allocation and deallocation patterns, such as sequential and
random memory accesses, to stress-test the allocator under controlled conditions.
Macro benchmarks involved real-world applications, offering insights into how
the allocator performs with complex memory allocation demands, large datasets,
and varying execution contexts.
The results section presents the outcomes of our benchmarks, highlighting key metrics
such as TLB miss rates, memory usage, and runtime performance. We observed that the
proposed allocator demonstrated significant improvements in reducing TLB misses,
leading to noticeable enhancements in runtime efficiency for both micro and macro
benchmarks. The behavior of specific allocation patterns and their impact on memory
performance is detailed, providing a nuanced understanding of the allocator's effectiveness.
Based on the evaluated results, the usability of the proposed allocator shows promise
for applications requiring optimized memory management and reduced overhead from TLB misses.
However, limitations were also identified, such as scenarios where the allocator's performance
gains were marginal or where it introduced additional complexity in memory management. These
limitations provide a roadmap for future optimizations and refinements of the allocator design.
** Expirement setup
#+BEGIN_COMMENT
The CHERI morello board\cite{Morello} was used to evaluate tehe proposed memory allocator.
Morello implements the ARM A76 with enhanced server class memory. The speciafication
includes a quad core ARM CPU with capabilties. The L1 and L2 cache was modified to
proliferate the capability bit. When compiling the C program for benchmarking
the Benchmark ABI\cite{BenchmarkABI} was used as recommended by the CHERI community as a compliation
mode with the Clang compilier.
The benchmark ABI was designed because the Morello
branch-predictor was not expanded to predict bounds. As a result, a capability-based
jump will stall later PCC-dependent instructions until bounds are established.
This is particularly problematic across dynamically linked calls
(and returns) between libraries, which will change bounds to those
covering the called (or returned-to) library.
Each C program is executed with 2 memory allocators. The first one being
the modified C allocator which is imported as a header file. This is because
the benchmark ABI shared object file has an unexpected behavoir of not overwriting
the C program on run time with the expected malloc functions to be overwritten.
The 2nd one being the standard OS memory allocator which in the case of CHERIBSD
is Jemalloc. The measurements are done using the ARM performance counters as mentioned
in the following section.
#+END_COMMENT
The CHERI Morello\cite{Morello} board was used to evaluate the proposed memory allocator.
Morello implements the ARM A76 with enhanced server-class memory, featuring a
quad-core ARM CPU with capability extensions. The L1 and L2 caches were modified
to proliferate the capability bit, ensuring compatibility with CHERI's capability-based
memory model. When compiling the C programs for benchmarking, the Benchmark ABI was
used as recommended by the CHERI community. This compilation mode was enabled using
the Clang compiler.
The Benchmark ABI\cite{BenchmarkABI} was specifically designed because the Morello branch predictor
was not expanded to predict bounds. Consequently, a capability-based jump introduces
stalls in later PCC-dependent instructions until bounds are established. This issue
is particularly significant during dynamically linked calls and returns between
libraries, where bounds are changed to cover the called or returned-to library.
Such stalls can negatively affect performance, making the Benchmark ABI an essential
consideration for this evaluation.
Each C program was executed using two different memory allocators. The first was
the modified C allocator, imported as a header file. This approach was necessary
because the Benchmark ABI shared object file exhibited unexpected behavior,
failing to overwrite the C program at runtime with the intended malloc functions.
The second allocator was the standard OS memory allocator, which, in the case of
CHERIBSD, is Jemalloc.
Performance measurements were carried out using ARM performance counters\cite{PerformanceCounter} to
ensure accurate evaluation. These counters provided detailed metrics, allowing
us to compare the performance of the two allocators and assess the impact of
the proposed changes.
#+CAPTION: ARM performance counters
#+NAME: fig:ARMPerformaneCounter
+--------------------------------------------------+--------------------------------------------+
| Performance counter | Description |
+--------------------------------------------------+--------------------------------------------+
| Wall clock | The actual time taken from the start of a |
| | computer program to the end. |
| | |
| (p/l1d_tlb_rd) L1 data TLB reads | Level 1 data TLB access, read |
| | |
| (p/l2d_tlb_rd) L2 data TLB reads | Level 2 data TLB access, read |
| | |
| (p/l1d_tlb_refill) L1 data TLB refills | Level 1 data TLB refill. |
| | The Level 1 data TLB refill |
| | counter tracks each access to |
| | the L1D_TLB that results |
| | in a refill of the Level 1 data |
| | or unified TLB. This includes any |
| | access that requires a memory lookup |
| | due to a translation table walk |
| | or accessing another level of TLB cache. |
| | |
| (p/cpu_cycles) CPU cycles | The CPU CYCLES counter increases with |
| | every clock cycle. However, it can be |
| | affected by changes in clock frequency, |
| | such as when WFI (Wait for Interrupt) |
| | or WFE (Wait for Event) |
| | instructions pause the clock. |
| | |
| (p/dtlb_walk) Data TLB walks | Data TLB access with at least |
| | one translation table walk. |
| | |
| (p/ll_cache_miss_rd) Last level cache miss reads | Last level cache miss, read |
| | (This refers to every miss in the |
| | Last level cache that occurs |
| | during a memory read operation.) |
+--------------------------------------------------+--------------------------------------------+
*** Benchmarks
The benchmarks\cite{Benchmark} are classified into 2 classes:
**** Micro benchmark
- GLIBC: The Glibc benchmark evaluates the performance of
malloc and free functions in single-threaded, multi-threaded,
and emulated multi-threading scenarios using various block sizes and
allocation patterns. It simulates real-world memory usage by partially
deallocating blocks in FIFO order and fully deallocating them in LIFO order.
Results are gathered across configurations to analyze performance variations.
- MemAccess: This benchmark by Alex Bordei evaluates the performance impact of
memory access patterns by constructing and traversing a doubly
linked list with varying working set sizes. It supports sequential or
randomized structures, optional node operations, and multithreaded
traversal using pthreads. The program dynamically allocates memory and systematically
doubles the working set size to analyze memory hierarchy behavior.
**** Macro runs
- Kmeans: Kmeans implements a parallelized K-means clustering algorithm that
assigns data points to clusters based on proximity to centroids,
iteratively updating them until convergence. The computation is
distributed across threads using the pthread library, dynamically
assigning tasks to optimize performance. Parameters like data size
and clusters are configurable, and the program ensures efficient
memory management and synchronization.
- Richards: Richards is a task scheduling benchmark that simulates a
multitasking environment with tasks of varying types and priorities,
communicating through queued packets. The schedule function manages
task execution based on state and priority, tracking processed packets
and held tasks for performance evaluation. Configurable iterations and
timing help measure system performance and ensure correctness.
** Results
#+ATTR_HTML: :align right
#+ATTR_ORG: :align center
#+CAPTION: Percentage difference between the modified memory allocator against the default system memory allocator
#+NAME: fig:bargraph
[[./diagrams/bargraph.png]]
#+BEGIN_COMMENT
The graph above refers to the precentage difference between the modified
memory allocator against the default system memory allocator which is
Jemalloc. Since FAT pointer memory allocator is desgined to allocate
with huge pages the results in graph above has the appripirate
expected corresponding behavoir. It is noticable the data
TLB walk, L2 data TLB reads and refill are consistently
90% lesser than the default memory allocator accross
the benchmarks listed on the graph above. This is
because of a single huge page entry at the l1 TLB
layer. This means most address translations hit L1
TLB without having to walk through the heirarchy of
TLB translations.
The micro benchmarks are designed for more memory reads
and shows on average a 50% reduction on wallclock runtimes.
The macro benchmarks on the other hand which are larger
classes of C programs have minimal differences in wall
clock run times.
#+END_COMMENT
The graph[[fig:bargraph]] highlights the performance comparison between the modified memory allocator and
Jemalloc, the default memory allocator. The FAT pointer memory allocator, specifically optimized
for use with huge pages, demonstrates a clear advantage in scenarios where memory allocation
patterns benefit from its design. The results align with expectations, showcasing the impact
of its capability to handle memory more efficiently by leveraging huge pages.
A particularly striking observation is the significant reduction in data TLB walks,
L2 data TLB reads, and TLB refills—consistently showing a 90% decrease across all
benchmarks compared to Jemalloc. This improvement is due to the modified allocator's
use of a single huge page entry at the L1 TLB layer. By enabling most address translations
to be resolved directly at the L1 TLB, the need to walk through the deeper TLB hierarchy is
largely eliminated. This reduction in translation overhead is a key factor in the allocator's
superior performance for certain types of workloads.
The micro benchmarks, which are crafted to emphasize memory read operations, highlight the
allocator's strengths. These tests simulate frequent and intensive memory access patterns,
where the reduction in TLB misses directly translates into measurable performance gains.
On average, the FAT pointer allocator achieves a 50% reduction in wall clock runtimes for
these workloads, underscoring its ability to optimize high-throughput memory operations.
On the other hand, macro benchmarks, which represent larger and more complex real-world applications,
exhibit minimal differences in wall clock runtimes when using the FAT pointer allocator.
This outcome is expected, as macro benchmarks typically involve a broader range of operations
beyond memory allocation, diluting the impact of the allocator's optimizations. Additionally,
the benefits of huge pages may be less pronounced for these workloads, as they are often
bottlenecked by factors such as computation or I/O rather than memory translation overhead.
#+ATTR_HTML: :align right
#+ATTR_ORG: :align center
#+CAPTION: Kmeans COZ benchmark executed against various cluster sizes
#+NAME: fig:Kmeans
[[./diagrams/kmeans.png]]
#+BEGIN_COMMENT
The kmeans was executed with various cluster sizes to see
the percentage difference against the baseline allocator as
the size of the workload increases. It can be noted that
the percentage difference stays the same except during
the cluster size of 2000.
#+END_COMMENT
The K-means algorithm was executed with varying cluster sizes to evaluate the performance difference
between the FAT pointer allocator and the baseline allocator as the workload scales. This analysis
aimed to understand how the allocator's optimizations, particularly its ability to manage memory
more efficiently with huge pages, impact performance under different workload conditions.
For most cluster sizes tested, the percentage difference in performance remained relatively
consistent. This indicates that the allocator's efficiency scales predictably with increasing
workload sizes, suggesting a stable and uniform benefit across different configurations. The
consistent performance gain is likely due to the allocator's ability to minimize TLB misses
and efficiently manage memory allocations for the centroid and data point structures used in
the K-means algorithm.
However, an anomaly was observed at a cluster size of 2000, where the percentage difference
deviated significantly from the trend. This irregularity could be attributed to several factors.
At this cluster size, the memory access patterns and allocation behavior may align in a way that
temporarily offsets the advantages of the FAT pointer allocator. For example, the memory layout
might interact with system-level caching mechanisms or TLB behavior differently, leading to an
unexpected change in performance. Additionally, the increased complexity of managing a higher
number of clusters might introduce computational overhead that overshadows the memory allocator's
optimizations.
This observation highlights the importance of testing across a range of workload sizes and
configurations to uncover edge cases or specific scenarios where performance deviates from the
expected pattern. Understanding these anomalies can provide insights into the allocator's
behavior and guide future improvements to address such outliers. Despite the deviation at a
cluster size of 2000, the overall results reaffirm the allocator's capability to maintain
consistent performance benefits across most scenarios.
#+BEGIN_COMMENT
#+ATTR_HTML: :align right
#+ATTR_ORG: :align center
[[./diagrams/glibc.png]]
#+END_COMMENT
** Usability
The FAT pointer memory allocator demonstrates significant potential for enhancing
memory management in systems that benefit from huge page optimizations. Its design
effectively reduces TLB misses, achieving up to 90% fewer data TLB walks, L2 TLB reads,
and TLB refills compared to Jemalloc. These improvements lead to noticeable performance
gains, especially in micro benchmarks, where the allocator reduces wall clock runtimes
by an average of 50%.
The allocator integrates seamlessly into memory-intensive workloads, as evidenced by its
consistent performance across varying cluster sizes in the K-means benchmark, with only
minor anomalies observed under specific conditions. These outliers provide valuable
insights into the allocator's interaction with system-level caching and memory translation mechanisms.
While the allocator excels in scenarios emphasizing high memory throughput, its impact on
macro benchmarks is less pronounced. This suggests that its benefits are most relevant for
applications with frequent and intensive memory operations rather than those constrained by
computation or I/O bottlenecks.
\bibliographystyle{IEEEtran}
\bibliography{evaluation.bib}