Disk I/O Limitations: When Storage Becomes the Bottleneck

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57

#!/bin/bash
# =====================================================
# STORAGE BENCHMARKING WITH fio
# =====================================================
 
# fio: Flexible I/O Tester - the standard for storage benchmarking
 
# Random Read IOPS (4KB blocks, 32 queue depth)
fio --name=randread --ioengine=libaio --iodepth=32 \
    --rw=randread --bs=4k --direct=1 --size=4G \
    --numjobs=4 --runtime=60 --group_reporting
 
# Random Write IOPS
fio --name=randwrite --ioengine=libaio --iodepth=32 \
    --rw=randwrite --bs=4k --direct=1 --size=4G \
    --numjobs=4 --runtime=60 --group_reporting
 
# Sequential Read Throughput
fio --name=seqread --ioengine=libaio --iodepth=32 \
    --rw=read --bs=1M --direct=1 --size=4G \
    --numjobs=1 --runtime=60 --group_reporting
 
# Sequential Write Throughput
fio --name=seqwrite --ioengine=libaio --iodepth=32 \
    --rw=write --bs=1M --direct=1 --size=4G \
    --numjobs=1 --runtime=60 --group_reporting
 
 
# =====================================================
# INTERPRETING fio OUTPUT
# =====================================================
 
# Sample output:
# read: IOPS=245k, BW=958MiB/s (1005MB/s)(56.2GiB/60001msec)
#    lat (usec): min=4, max=9876, avg=126.43, stdev=312.45
#
# Key metrics:
# - IOPS: 245,000 random reads/second
# - BW: 958 MB/s bandwidth
# - lat avg: 126 microseconds average latency
# - lat stdev: 312 microseconds (variability)
 
 
# =====================================================
# QUICK DISK PERFORMANCE CHECK
# =====================================================
 
# Write test (careful: writes to current directory)
dd if=/dev/zero of=testfile bs=1G count=1 oflag=direct
# Shows sequential write speed
 
# Read back
dd if=testfile of=/dev/null bs=1G count=1 iflag=direct
# Shows sequential read speed
 
# Clean up
rm testfile

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95

# =====================================================
# I/O ACCESS PATTERN COMPARISON
# =====================================================
 
import os
import time
 
# =====================================================
# ANTI-PATTERN: Many small synchronous writes
# =====================================================
 
def write_records_bad(records: list, filepath: str):
    """Write each record separately - terrible I/O performance."""
    for record in records:
        with open(filepath, 'ab') as f:  # Open/close per record!
            f.write(record.encode() + b'
')
            f.flush()  # Force to disk
            os.fsync(f.fileno())  # Wait for disk
 
# 10,000 records at 3 IOPS (HDD) = 55 minutes!
# 10,000 records at 100K IOPS (SSD) = 0.1 seconds
# But we're also opening/closing file 10,000 times!
 
 
# PATTERN: Batch writes with proper buffering
def write_records_good(records: list, filepath: str):
    """Batch records and write efficiently."""
    with open(filepath, 'ab') as f:
        for record in records:
            f.write(record.encode() + b'
')
        f.flush()  # One flush for all records
        os.fsync(f.fileno())  # One sync
 
# 10,000 records = 1 I/O operation (plus OS buffering)
 
 
# =====================================================
# ANTI-PATTERN: Sequential reads of random positions
# =====================================================
 
def process_ids_bad(ids: list, data_file: str):
    """Read each ID's data in order given - random access."""
    results = []
    for id in ids:  # IDs are not sorted by file position
        offset = get_offset_for_id(id)
        with open(data_file, 'rb') as f:
            f.seek(offset)  # Random seek
            data = f.read(1024)
            results.append(process(data))
    return results
 
# 10,000 IDs with random seeks on HDD = ~100 seconds
# Each seek is ~10ms
 
 
# PATTERN: Sort reads by position, then process
def process_ids_good(ids: list, data_file: str):
    """Sort IDs by file position for sequential access."""
    # Get offsets and sort
    id_offsets = [(id, get_offset_for_id(id)) for id in ids]
    id_offsets.sort(key=lambda x: x[1])  # Sort by offset
    
    results = []
    with open(data_file, 'rb') as f:
        for id, offset in id_offsets:
            f.seek(offset)
            data = f.read(1024)
            results.append((id, process(data)))
    
    # Reorder results back to original ID order
    return [r for id, r in sorted(results, key=lambda x: ids.index(x[0]))]
 
# Sequential access: 10,000 reads in ~1 second on HDD
 
 
# =====================================================
# PATTERN: Async I/O for parallel operations
# =====================================================
 
import asyncio
import aiofiles
 
async def read_files_parallel(filepaths: list):
    """Read multiple files in parallel - maximizes I/O utilization."""
    async def read_one(path):
        async with aiofiles.open(path, 'r') as f:
            return await f.read()
    
    return await asyncio.gather(*[read_one(p) for p in filepaths])
 
# 100 files with 10ms latency each:
# Sequential: 100 * 10ms = 1 second
# Parallel: ~10ms (all in flight simultaneously)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99

#!/bin/bash
# =====================================================
# STEP 1: CHECK FOR I/O WAIT
# =====================================================
 
# Is CPU waiting on I/O?
top
# Look at '%wa' (I/O wait) in the CPU line
# High %wa (>10%) indicates disk bottleneck
 
# Or use vmstat
vmstat 1
# 'wa' column shows I/O wait percentage
 
# sar provides history
sar -u 1 10
# %iowait column
 
 
# =====================================================
# STEP 2: IDENTIFY BUSY DISKS
# =====================================================
 
iostat -xz 1
# Output columns explained:
# rrqm/s, wrqm/s - read/write merges (kernel coalescing)
# r/s, w/s       - reads/writes per second
# rMB/s, wMB/s   - throughput
# await          - average I/O wait time (ms) - KEY METRIC
# %util          - percentage of time disk was busy - KEY METRIC
 
# Example output:
# Device   r/s    w/s   rMB/s  wMB/s  await  %util
# sda    245.00  15.00   1.23   0.12  24.35  78.50
 
# Interpretation:
# - 245 reads/second, 15 writes/second
# - await 24ms is HIGH for SSD (should be <1ms)
# - %util 78% means disk is busy 78% of the time
# - %util >70% sustained = bottleneck
 
 
# =====================================================
# STEP 3: IDENTIFY I/O-HEAVY PROCESSES
# =====================================================
 
# Real-time I/O per process
iotop -o  # Show only processes with I/O
 
# Output:
# TID  PRIO  USER     DISK READ  DISK WRITE  SWAPIN     IO>    COMMAND
# 1234 be/4  mysql     12.45 M/s   3.21 M/s  0.00 %  87.23 %  mysqld
 
# IO> column shows percentage of time this process is waiting on I/O
 
 
# Process-level I/O stats
pidstat -d 1
# Shows kB_rd/s (read), kB_wr/s (write) per process
 
 
# =====================================================
# STEP 4: ANALYZE I/O PATTERNS
# =====================================================
 
# What operations are happening?
# Using blktrace for block-level I/O tracing
 
blktrace -d /dev/sda -o trace
# Run workload...
blkparse -i trace.blktrace.* -o trace.txt
 
# Or use bcc/eBPF tools (modern Linux)
# biolatency - I/O latency histogram
/usr/share/bcc/tools/biolatency
# Shows distribution of I/O latency
 
# biosnoop - trace each I/O with latency
/usr/share/bcc/tools/biosnoop
# Real-time trace of every I/O operation
 
 
# =====================================================
# STEP 5: AWS EBS-SPECIFIC DIAGNOSIS
# =====================================================
 
# EBS has IOPS and throughput limits
# Check CloudWatch metrics:
# - VolumeReadOps, VolumeWriteOps (total ops)
# - VolumeQueueLength (should be <1 ideally)
# - BurstBalance (for gp2/gp3, shows burst credit remaining)
 
# If BurstBalance is 0: you've exhausted IOPS credits
# If VolumeQueueLength >10: I/O is queuing up
 
# Check instance EBS bandwidth:
# Some instances have shared EBS bandwidth
# c5.xlarge: baseline 4750 Mbps EBS bandwidth
# If EBS usage exceeds this, you're throttled

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73

-- =====================================================
-- POSTGRESQL I/O OPTIMIZATION
-- =====================================================
 
-- Check cache hit ratio (should be >99% for OLTP)
SELECT 
    sum(heap_blks_read) as heap_read,
    sum(heap_blks_hit) as heap_hit,
    sum(heap_blks_hit) / (sum(heap_blks_hit) + sum(heap_blks_read))::float as cache_hit_ratio
FROM pg_statio_user_tables;
 
-- Low cache_hit_ratio (<95%) = data doesn't fit in memory
-- Solution: increase shared_buffers or reduce working set
 
 
-- Check index usage
SELECT 
    relname,
    idx_scan as index_scans,
    seq_scan as sequential_scans,
    n_tup_upd as updates,
    n_tup_del as deletes
FROM pg_stat_user_tables
WHERE seq_scan > 0
ORDER BY seq_scan DESC;
 
-- High sequential scans on large tables = missing indexes
-- Each seq_scan reads entire table from disk
 
 
-- I/O-heavy queries (pg_stat_statements extension)
SELECT query, 
       calls,
       total_time,
       blk_read_time,  -- Time spent reading from disk
       blk_write_time, -- Time spent writing to disk
       (blk_read_time + blk_write_time) / total_time * 100 as io_percent
FROM pg_stat_statements
ORDER BY blk_read_time DESC
LIMIT 10;
 
-- High io_percent queries are I/O-bound
-- Candidates for optimization: indexes, caching
 
 
-- =====================================================
-- KEY POSTGRESQL I/O SETTINGS
-- =====================================================
 
-- shared_buffers: PostgreSQL's buffer cache
-- Typically 25% of RAM for dedicated DB server
-- Example: 4GB on a 16GB server
ALTER SYSTEM SET shared_buffers = '4GB';
 
-- effective_cache_size: Tell planner about OS cache
-- Typically 50-75% of RAM
ALTER SYSTEM SET effective_cache_size = '12GB';
 
-- work_mem: Memory for sorting/hashing per operation
-- Per-operation! 10 concurrent sorts × 64MB = 640MB
ALTER SYSTEM SET work_mem = '64MB';
 
-- maintenance_work_mem: Memory for VACUUM, CREATE INDEX
ALTER SYSTEM SET maintenance_work_mem = '1GB';
 
-- wal_buffers: WAL buffer size (16-64MB typically)
ALTER SYSTEM SET wal_buffers = '64MB';
 
-- checkpoint_timeout, max_wal_size: Control checkpoint frequency
-- More frequent = more consistent I/O, more total I/O
-- Less frequent = less total I/O, spikier performance
ALTER SYSTEM SET checkpoint_timeout = '30min';
ALTER SYSTEM SET max_wal_size = '4GB';

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151

# =====================================================
# BUFFERED WRITING PATTERN
# =====================================================
 
import threading
import time
from queue import Queue
from typing import Any
 
class BufferedWriter:
    """Batch writes to reduce I/O operations."""
    
    def __init__(self, filepath: str, buffer_size: int = 1000,
                 flush_interval: float = 1.0):
        self.filepath = filepath
        self.buffer_size = buffer_size
        self.flush_interval = flush_interval
        self.buffer = []
        self.lock = threading.Lock()
        
        # Background flush thread
        self._flush_thread = threading.Thread(target=self._periodic_flush)
        self._flush_thread.daemon = True
        self._flush_thread.start()
    
    def write(self, data: str):
        """Add data to buffer; non-blocking."""
        with self.lock:
            self.buffer.append(data)
            if len(self.buffer) >= self.buffer_size:
                self._flush()
    
    def _flush(self):
        """Write buffer to disk."""
        if not self.buffer:
            return
        
        # Take ownership of buffer
        to_write = self.buffer
        self.buffer = []
        
        # Single I/O operation for entire batch
        with open(self.filepath, 'a') as f:
            f.writelines(to_write)
    
    def _periodic_flush(self):
        """Ensure data is written even if buffer isn't full."""
        while True:
            time.sleep(self.flush_interval)
            with self.lock:
                self._flush()
 
# Usage:
writer = BufferedWriter('/var/log/app.log', buffer_size=1000)
for event in events:
    writer.write(event)  # Fast, buffers in memory
 
# Result: 1000 individual writes become 1 disk write
 
 
# =====================================================
# ASYNC FILE OPERATIONS
# =====================================================
 
import asyncio
import aiofiles
 
async def process_files_async(filepaths: list):
    """Process multiple files with non-blocking I/O."""
    
    async def process_one(filepath):
        async with aiofiles.open(filepath, 'r') as f:
            content = await f.read()
            # Process content...
            return len(content)
    
    # All files read in parallel - overlapping I/O
    results = await asyncio.gather(*[process_one(p) for p in filepaths])
    return results
 
# 100 files × 10ms each:
# Sequential: 1000ms
# Parallel: ~10-20ms (limited by concurrent I/O capacity)
 
 
# =====================================================
# MEMORY-MAPPED FILES FOR READ-HEAVY WORKLOADS
# =====================================================
 
import mmap
 
class MmapReader:
    """Memory-mapped file access - let OS manage caching."""
    
    def __init__(self, filepath: str):
        self.file = open(filepath, 'rb')
        self.mmap = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ)
    
    def read_at(self, offset: int, length: int) -> bytes:
        """Read from any position - OS caches in page cache."""
        return self.mmap[offset:offset + length]
    
    def close(self):
        self.mmap.close()
        self.file.close()
 
# Advantages:
# - OS handles caching automatically
# - Zero-copy access to cached data
# - Efficient for random access patterns
# - Memory usage = what's currently accessed, not whole file
 
 
# =====================================================
# WRITE-AHEAD LOG PATTERN
# =====================================================
 
class WALWriter:
    """
    Write-ahead log: convert random writes to sequential.
    Used by databases, message queues, and other durable systems.
    """
    
    def __init__(self, log_path: str):
        self.log_file = open(log_path, 'ab')
    
    def append(self, operation: dict) -> int:
        """Append operation to log; return log sequence number."""
        import json
        
        # Serialize
        data = json.dumps(operation).encode() + b'
'
        
        # Sequential append (very fast)
        position = self.log_file.tell()
        self.log_file.write(data)
        self.log_file.flush()
        
        return position
    
    def sync(self):
        """Ensure log is durable on disk."""
        import os
        os.fsync(self.log_file.fileno())
 
# Pattern:
# 1. Write operation to WAL (sequential, fast)
# 2. Apply operation to data structures (in memory)
# 3. Periodically checkpoint data to disk (batch random writes)
# 4. Truncate old WAL entries after checkpoint

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

#!/bin/bash
# =====================================================
# CLOUD STORAGE PERFORMANCE CHECKLIST (AWS)
# =====================================================
 
# 1. Check instance EBS bandwidth
# Look up instance type in AWS docs
# Example: m5.xlarge = 4,750 Mbps EBS bandwidth
 
# 2. Check volume configuration
aws ec2 describe-volumes --volume-id vol-xxxxx
# Verify: VolumeType, Size, Iops, Throughput
 
# 3. Monitor CloudWatch metrics
# VolumeReadOps, VolumeWriteOps - total ops
# VolumeReadBytes, VolumeWriteBytes - throughput
# VolumeQueueLength - pending I/O (should be <1)
# BurstBalance - burst credits remaining (gp2)
 
# 4. Check for EBS throughput throttling
# CloudWatch: EBSIOBalance%, EBSByteBalance%
# If these drop to 0, you're being throttled
 
 
# =====================================================
# RECOMMENDATIONS BY WORKLOAD
# =====================================================
 
# Workload: Database (high IOPS random)
# - Use io2 for production databases
# - Provision IOPS based on peak, not average
# - Monitor VolumeQueueLength for congestion
 
# Workload: Analytics (high throughput sequential)
# - gp3 with high throughput provisioned
# - Or st1 for cold data with sequential access
# - Use multiple volumes in RAID 0 for higher throughput
 
# Workload: Logs (sequential write, rare read)
# - gp3 baseline (3000 IOPS, 125 MB/s)
# - Or st1 for very high volume, low performance need
 
# Workload: General application
# - Start with gp3 (better $/IOPS than gp2)
# - Provision IOPS/throughput as needed
# - Monitor and adjust
 
 
# =====================================================
# GCP PERSISTENT DISK COMPARISON
# =====================================================
 
# pd-standard: HDD, cheapest, low performance
# pd-balanced: SSD, moderate cost/performance
# pd-ssd: SSD, high performance
# pd-extreme: Up to 120K IOPS
 
# GCP disks auto-scale IOPS with size (unlike AWS)
# 30 IOPS/GB for pd-ssd, up to 100K IOPS max