Add write buffering to reduce IOPS from small sequential writes

.gitignore

@@ -30,3 +30,10 @@
 *.exe
 *.out
 *.app
+
+# Build directories
+build/
+CMakeFiles/
+CMakeCache.txt
+cmake_install.cmake
+Makefile

README.md

@@ -25,6 +25,26 @@ To enable the checksum (only on XRootD 5.2+):
 ofs.ckslib * libXrdMultiuser.so
 ```
 
+### Write Buffering
+
+To reduce IOPS from small sequential writes, you can enable write buffering:
+
+```
+multiuser.writebuffersize <bytes>
+```
+
+Where `<bytes>` is the buffer size in bytes. Default is 0 (disabled). When enabled:
+- Sequential writes smaller than the buffer size are accumulated in memory
+- The buffer is flushed when full, when a non-sequential write occurs, or when the file is closed
+- Buffering is automatically disabled for a file if non-sequential writes are detected
+
+Example: Buffer up to 1MB of writes:
+```
+multiuser.writebuffersize 1048576
+```
+
+**Note:** Buffering is only suitable for sequential write workloads. Non-sequential writes will cause the buffer to be flushed and buffering disabled for that file.
+
 Startup
 -------
 

configs/60-osg-multiuser.cfg

@@ -17,4 +17,9 @@ if defined ?~XC_ENABLE_MULTIUSER
   # checksum while it is writing a file.  To turn this on, uncomment the
   # following line:
   # multiuser.checksumonwrite on
+
+  # Write buffering can reduce IOPS for workloads with many small sequential
+  # writes. Specify the buffer size in bytes (default: 0 = disabled).
+  # Example: Buffer up to 1MB of sequential writes
+  # multiuser.writebuffersize 1048576
 fi

src/MultiuserFile.hh

@@ -15,10 +15,12 @@
 #include "XrdChecksum.hh"
 
 #include <memory>
+#include <mutex>
+#include <vector>
 
 class MultiuserFile : public XrdOssDF {
 public:
-    MultiuserFile(const char *user, std::unique_ptr<XrdOssDF> ossDF, XrdSysError &log, mode_t umask_mode, bool checksum_on_write, unsigned digests, MultiuserFileSystem *oss);
+    MultiuserFile(const char *user, std::unique_ptr<XrdOssDF> ossDF, XrdSysError &log, mode_t umask_mode, bool checksum_on_write, unsigned digests, MultiuserFileSystem *oss, size_t write_buffer_size);
 
     virtual ~MultiuserFile() {
             if (m_state) {delete m_state;}
@@ -127,6 +129,7 @@ public:
     int Close(long long *retsz=0);
 
 private:
+    int FlushWriteBuffer();
     std::unique_ptr<XrdOssDF> m_wrapped;
     XrdSysError &m_log;
     const XrdSecEntity* m_client;
@@ -138,6 +141,13 @@ private:
     bool m_checksum_on_write;
     unsigned m_digests;
 
+    // Write buffering
+    size_t m_write_buffer_size;
+    std::vector<unsigned char> m_write_buffer;
+    off_t m_buffer_offset;
+    bool m_buffering_enabled;
+    std::mutex m_buffer_mutex;
+
 };
 
 #endif

src/MultiuserFileSystem.cc

@@ -41,7 +41,8 @@ MultiuserFileSystem::MultiuserFileSystem(XrdOss *oss, XrdSysLogger *lp, const ch
     m_env(envP),
     m_log(lp, "multiuser_"),
     m_checksum_on_write(false),
-    m_digests(0)
+    m_digests(0),
+    m_write_buffer_size(0)
 {
     if (!oss) {
         throw std::runtime_error("The multi-user plugin must be chained with another filesystem.");
@@ -114,6 +115,28 @@ MultiuserFileSystem::Config(XrdSysLogger *lp, const char *configfn)
                 return false;
             }
         }
+
+        // Write buffer size
+        if (!strcmp("multiuser.writebuffersize", val)) {
+            val = Config.GetWord();
+            if (!val || !val[0]) {
+                m_log.Emsg("Config", "multiuser.writebuffersize must specify a value");
+                Config.Close();
+                return false;
+            }
+            char *endptr = NULL;
+            errno = 0;
+            long int buffer_size = strtol(val, &endptr, 0);
+            if (errno == ERANGE || buffer_size < 0 || endptr == val || *endptr != '\0') {
+                m_log.Emsg("Config", "multiuser.writebuffersize must specify a valid non-negative integer");
+                Config.Close();
+                return false;
+            }
+            m_write_buffer_size = static_cast<size_t>(buffer_size);
+            std::stringstream ss;
+            ss << "Setting write buffer size to " << m_write_buffer_size << " bytes";
+            m_log.Emsg("Config", ss.str().c_str());
+        }
         if (!strcmp("xrootd.chksum", val)) {
             m_digests = 0;
             val = Config.GetWord();
@@ -175,7 +198,7 @@ XrdOssDF *MultiuserFileSystem::newFile(const char *user)
 {
     // Call the underlying OSS newFile
     std::unique_ptr<XrdOssDF> wrapped(m_oss->newFile(user));
-    return (MultiuserFile *)new MultiuserFile(user, std::move(wrapped), m_log, m_umask_mode, m_checksum_on_write, m_digests, this);
+    return (MultiuserFile *)new MultiuserFile(user, std::move(wrapped), m_log, m_umask_mode, m_checksum_on_write, m_digests, this, m_write_buffer_size);
 }
 
 int MultiuserFileSystem::Chmod(const char * path, mode_t mode, XrdOucEnv *env)

src/MultiuserFileSystem.hh

@@ -24,6 +24,8 @@ public:
     bool
     Config(XrdSysLogger *lp, const char *configfn);
 
+    size_t GetWriteBufferSize() const { return m_write_buffer_size; }
+
     XrdOssDF *newDir(const char *user=0);
     XrdOssDF *newFile(const char *user=0);
     int Chmod(const char * path, mode_t mode, XrdOucEnv *env=0);
@@ -71,6 +73,7 @@ private:
     std::shared_ptr<XrdAccAuthorize> m_authz;
     bool m_checksum_on_write;
     unsigned m_digests;
+    size_t m_write_buffer_size;
 
 };
 

src/multiuser.cpp

@@ -133,7 +133,7 @@ class ErrorSentry
 };
 
 
-MultiuserFile::MultiuserFile(const char *user, std::unique_ptr<XrdOssDF> ossDF, XrdSysError &log, mode_t umask_mode, bool checksum_on_write, unsigned digests, MultiuserFileSystem *oss) :
+MultiuserFile::MultiuserFile(const char *user, std::unique_ptr<XrdOssDF> ossDF, XrdSysError &log, mode_t umask_mode, bool checksum_on_write, unsigned digests, MultiuserFileSystem *oss, size_t write_buffer_size) :
     XrdOssDF(user),
     m_wrapped(std::move(ossDF)),
     m_log(log),
@@ -142,7 +142,10 @@ MultiuserFile::MultiuserFile(const char *user, std::unique_ptr<XrdOssDF> ossDF,
     m_nextoff(0),
     m_oss(oss),
     m_checksum_on_write(checksum_on_write),
-    m_digests(digests)
+    m_digests(digests),
+    m_write_buffer_size(write_buffer_size),
+    m_buffer_offset(-1),
+    m_buffering_enabled(write_buffer_size > 0)
 {}
 
 int     MultiuserFile::Open(const char *path, int Oflag, mode_t Mode, XrdOucEnv &env)
@@ -173,28 +176,143 @@ int     MultiuserFile::Open(const char *path, int Oflag, mode_t Mode, XrdOucEnv
 
 ssize_t MultiuserFile::Write(const void *buffer, off_t offset, size_t size)
 {
+    // Lock protects buffer state and sequential write tracking.
+    // While this serializes writes to the same file, it ensures correctness
+    // and is consistent with the sequential write assumption.
+    std::lock_guard<std::mutex> lock(m_buffer_mutex);
 
-    if ((offset != m_nextoff) && m_state) 
+    // Check for out-of-order writes if checksumming or buffering
+    if ((offset != m_nextoff) && (m_state || m_buffering_enabled)) 
     {   
-        std::stringstream ss;
-        ss << "Out-of-order writes not supported while running checksum. " << m_fname;
-        m_log.Emsg("Write", ss.str().c_str());
-        return -ENOTSUP;
+        // Flush any buffered data first
+        if (m_buffering_enabled) {
+            int flush_result = FlushWriteBuffer();
+            if (flush_result < 0) {
+                return flush_result;
+            }
+            // Disable buffering for the rest of this file
+            m_buffering_enabled = false;
+        }
+
+        if (m_state) {
+            std::stringstream ss;
+            ss << "Out-of-order writes not supported while running checksum. " << m_fname;
+            m_log.Emsg("Write", ss.str().c_str());
+            return -ENOTSUP;
+        }
     }
 
+    // If buffering is enabled and configured
+    if (m_buffering_enabled) {
+        // If this is the first write or buffer is empty, initialize buffer offset
+        if (m_write_buffer.empty()) {
+            m_buffer_offset = offset;
+        }
+
+        // Check if this write is sequential to the buffer
+        off_t expected_offset = m_buffer_offset + static_cast<off_t>(m_write_buffer.size());
+        if (offset != expected_offset) {
+            // Not sequential - flush buffer and disable buffering
+            int flush_result = FlushWriteBuffer();
+            if (flush_result < 0) {
+                return flush_result;
+            }
+            m_buffering_enabled = false;
+            // Fall through to direct write
+        } else {
+            // Sequential write - check if we should buffer it
+            size_t total_size = m_write_buffer.size() + size;
+
+            if (total_size <= m_write_buffer_size) {
+                // Buffer this write - reserve capacity to avoid reallocations
+                if (m_write_buffer.capacity() < total_size) {
+                    m_write_buffer.reserve(total_size);
+                }
+                m_write_buffer.insert(m_write_buffer.end(), 
+                                     static_cast<const unsigned char*>(buffer), 
+                                     static_cast<const unsigned char*>(buffer) + size);
+                m_nextoff = offset + size;
+                return size;
+            } else {
+                // Buffer would exceed limit - flush and write directly
+                int flush_result = FlushWriteBuffer();
+                if (flush_result < 0) {
+                    return flush_result;
+                }
+                // Fall through to direct write
+            }
+        }
+    }
+
+    // Direct write (no buffering or buffer disabled)
     auto result = m_wrapped->Write(buffer, offset, size);
-    if (result >= 0) {m_nextoff += result;}
-    if (m_state)
-    {
-        m_state->Update(static_cast<const unsigned char*>(buffer), size);
+    if (result >= 0) {
+        m_nextoff = offset + result;
+        if (m_state && result > 0) {
+            // Only update checksum for the data that was actually written
+            m_state->Update(static_cast<const unsigned char*>(buffer), result);
+        }
     }
     return result;
 }
 
 
 
+// FlushWriteBuffer: Writes all buffered data to the underlying file system.
+// Preconditions: Must be called with m_buffer_mutex held.
+// Returns: 0 on success, negative error code on failure.
+// Side effects:
+//   - Writes buffer contents via m_wrapped->Write()
+//   - Updates checksum state if enabled (m_state)
+//   - Clears m_write_buffer and resets m_buffer_offset on success
+//   - Handles partial writes with retry loop
+//   - On failure, buffer is NOT cleared to allow retry
+int MultiuserFile::FlushWriteBuffer()
+{
+    if (m_write_buffer.empty()) {
+        return 0;
+    }
+
+    size_t total_written = 0;
+    while (total_written < m_write_buffer.size()) {
+        auto result = m_wrapped->Write(m_write_buffer.data() + total_written, 
+                                       m_buffer_offset + total_written, 
+                                       m_write_buffer.size() - total_written);
+        if (result < 0) {
+            // Write failed - don't clear buffer, return error
+            return result;
+        }
+        if (result == 0) {
+            // No progress - treat as error
+            return -EIO;
+        }
+        total_written += result;
+    }
+
+    if (m_state) {
+        m_state->Update(m_write_buffer.data(), m_write_buffer.size());
+    }
+
+    m_write_buffer.clear();
+    m_buffer_offset = -1;
+
+    return 0;
+}
+
+
 int MultiuserFile::Close(long long *retsz) 
 {
+    std::lock_guard<std::mutex> lock(m_buffer_mutex);
+
+    // Flush any remaining buffered data
+    if (!m_write_buffer.empty()) {
+        int flush_result = FlushWriteBuffer();
+        if (flush_result < 0) {
+            m_log.Emsg("Close", "Failed to flush write buffer");
+            // Continue with close anyway
+        }
+    }
+
     auto close_result = m_wrapped->Close(retsz);
     if (m_state)
     {