In the vast landscape of the memory hierarchy, two tiers dominate the database professional's daily concerns: main memory (RAM) and persistent storage (disk). While caches accelerate and tapes archive, it is the boundary between RAM and disk that determines whether your database runs at blazing speed or grinding slowness.

This divide isn't merely a matter of degree—it represents fundamentally different technologies with fundamentally different behaviors. RAM stores data as electrical charges in capacitors, delivering nanosecond access with no mechanical movement. Disks—whether spinning magnetic platters or solid-state flash—store data in persistent physical states, requiring microseconds to milliseconds for each access.

Understanding the detailed characteristics of each technology is essential because every database optimization ultimately reduces to one question: How do we minimize the expensive crossings between these two worlds?

Main memory, commonly called RAM (Random Access Memory), is the primary working space for active database operations. When queries execute, tables are scanned, or indexes are traversed, the data flows through RAM. Understanding RAM's architecture explains both its remarkable speed and its fundamental limitation—volatility.

The DRAM Cell:

Modern RAM uses Dynamic RAM (DRAM) technology. Each bit is stored in a tiny structure consisting of:

• One transistor: Acts as an access gate
• One capacitor: Stores the actual data as electrical charge

This simplicity—just two components per bit—enables incredible density. Modern DRAM chips pack billions of cells onto a single die.

Why "Dynamic"?

The capacitor holding each bit leaks charge over time. Within milliseconds, the stored charge would dissipate, losing the data. To prevent this, DRAM controllers continuously refresh each cell, reading and rewriting its value before the charge decays. This refresh cycle consumes time and power—a tax for DRAM's density advantage over static alternatives.

DRAM Organization:

DRAM is organized hierarchically:

1. Cells: Individual bits arranged in rows and columns
2. Banks: Independent arrays that can operate in parallel
3. Ranks: Groups of chips accessed together to match bus width
4. Channels: Separate memory controllers providing bandwidth
5. DIMMs: Physical modules containing multiple chips

Accessing DRAM:

A memory access follows this sequence:

1. Row Activation (RAS): The memory controller selects a row. The entire row (typically 8KB) is loaded into the row buffer—a faster SRAM structure.
2. Column Access (CAS): Within the activated row, specific columns are read or written. If the desired data is in the already-active row, access is fast.
3. Precharge: Before accessing a different row in the same bank, the current row must be written back and the bank reset.

Row Buffer Hit vs. Miss:

• Row buffer hit: Requested address is in the currently active row → ~10ns
• Row buffer miss (closed): Different row needed, bank is closed → ~50-60ns
• Row buffer miss (conflict): Different row needed, must precharge first → ~70-80ns

These patterns explain why sequential access in RAM is faster than random access—sequential reads hit the same row buffer repeatedly.

Despite the rise of solid-state storage, hard disk drives remain crucial in database environments—particularly for bulk storage, backups, and archival systems where cost-per-gigabyte dominates over speed.

Physical Architecture:

An HDD is an electromechanical device consisting of:

1. Platters: Circular disks coated with magnetic material, spinning at high speed (5,400 - 15,000 RPM)
2. Read/Write Heads: Electromagnetic devices floating nanometers above the platter surface
3. Actuator Arm: Mechanical arm that positions the heads radially across the platter
4. Spindle Motor: Rotates the platters at constant speed
5. Controller Board: Electronics managing the interface and head positioning

Data Organization on Disk:

• Tracks: Concentric circles on a platter surface where data is recorded
• Sectors: Arc segments of a track, typically 512 bytes or 4KB (Advanced Format)
• Cylinders: The set of tracks at the same radial position across all platters
• Blocks: Logical groupings of sectors (often 4KB-64KB for databases)

The Three Components of Access Time:

Every random HDD access involves three delays:

1. Seek Time: Time to move the head assembly to the correct track

   • Modern drives: 4-12ms average
   • Worst case (full stroke): 15-20ms
   • Best case (adjacent track): <1ms

2. Rotational Latency: Time for the desired sector to rotate under the head

   • Average: Half a rotation (4.17ms for 7,200 RPM)
   • Best case: Sector happens to be there (0ms)
   • Worst case: Just missed, full rotation

3. Transfer Time: Time to read/write the data

   • Depends on data density and rotational speed
   • Modern drives: 100-200 MB/s sustained

Total random access time = Seek + Rotational Latency + Transfer ≈ 8-15ms

Sequential vs. Random Performance:

The sequential vs. random dichotomy is extreme for HDDs:

The 100x Gap: HDDs can read sequentially at 200 MB/s but only manage ~1 MB/s for random small reads. This 100x or greater gap fundamentally shapes how databases organize data on disk.

Database Implications:

• Table data files should be read sequentially when possible (full scans)
• Index structures (B-trees) minimize random disk accesses
• Transaction logs exploit sequential write patterns
• Buffer pools cache to avoid random reads
• Batch operations amortize seek time over multiple blocks

Solid-state drives have transformed database performance by eliminating the mechanical constraints of HDDs. With no moving parts, SSDs offer dramatically lower latency and vastly improved random access performance.

Flash Memory Fundamentals:

SSDs store data in NAND flash memory—a non-volatile storage technology based on floating-gate transistors. Each cell stores charge in a floating gate insulated from the control circuit, where it can persist without power indefinitely.

Cell Types (SLC, MLC, TLC, QLC):

NAND cells can store multiple bits by distinguishing voltage levels:

NAND Organization:

Flash is organized hierarchically:

1. Cells: Individual storage units (1-4 bits each)
2. Pages: Smallest read/write unit (4KB-16KB)
3. Blocks: Collection of pages (128-512 pages per block, ~512KB-4MB)
4. Planes: Independent block arrays for parallelism
5. Dies: Complete NAND chips
6. Packages: Multiple dies in one physical package

The Asymmetric Operations Problem:

NAND flash has a critical asymmetry:

• Read: Can read any page independently (~25-75μs)
• Write: Can only write to erased pages (~200-500μs)
• Erase: Can only erase entire blocks (~2-5ms)

This means you cannot simply overwrite data in place. To modify a page:

1. Read the entire block
2. Erase the block (slow)
3. Write all pages including the modified one

Flash Translation Layer (FTL):

To hide this complexity, SSDs implement a Flash Translation Layer:

• Maintains a logical-to-physical address mapping
• Redirects writes to free pages (out-of-place writes)
• Performs garbage collection to reclaim invalidated pages
• Implements wear leveling to distribute writes evenly

The FTL makes the SSD appear as a simple block device to the operating system and database.

SSD Performance Characteristics:

Interface Matters:

• SATA SSD: Limited to ~550 MB/s by the SATA interface (designed for HDDs)
• NVMe SSD: Uses PCIe lanes directly, achieving 3-7+ GB/s
• NVMe with multiple queues: 64K queues with 64K commands each—massive parallelism

Database Implications:

• Random read performance makes point queries viable without extensive caching
• Write patterns should still be optimized to reduce write amplification
• Log files benefit enormously from SSD speed
• Mixed workloads (OLTP) see the biggest gains
• Cold storage can remain on HDD for cost efficiency

Let's consolidate the performance differences between the three primary storage technologies that databases interact with. These numbers represent typical values for modern enterprise-grade components.

*DWPD = Drive Writes Per Day (the full drive capacity can be written this many times per day over the warranty period)

Key Ratios to Remember:

Different database operations exhibit different access patterns, and these patterns interact differently with each storage technology. Understanding these interactions guides storage allocation and query optimization.

Common Database Access Patterns:

Given the enormous performance gap between RAM and disk (even SSDs), databases employ sophisticated techniques to minimize the impact of slow storage access. These techniques are central to database architecture.

The Core Strategies:

Buffer Pool Deep Dive:

The buffer pool deserves special attention as the primary bridge between RAM and disk:

+-------------------+     +-------------------+
|   Query Executor  |     |   Storage Engine  |
+-------------------+     +-------------------+
          |                         |
          v                         v
+-------------------------------------------+
|              BUFFER POOL (RAM)            |
|   +-------+  +-------+  +-------+  ...    |
|   | Page  |  | Page  |  | Page  |         |
|   | (8KB) |  | (8KB) |  | (8KB) |         |
|   +-------+  +-------+  +-------+         |
|   [Clean]    [Dirty]    [Pinned]          |
+-------------------------------------------+
          |                         ^
          v                         |
+-------------------------------------------+
|         PERSISTENT STORAGE (Disk)         |
+-------------------------------------------+

Page States:

• Clean: Matches disk, safe to evict without writing
• Dirty: Modified in memory, must be written before eviction
• Pinned: Currently in use, cannot be evicted

Buffer Pool Operations:

1. Get Page: Check if in pool. If yes, return from RAM. If no, read from disk, add to pool, return.
2. Modify Page: Get page, modify in place, mark dirty.
3. Evict Page: Choose victim (LRU/clock). If dirty, write to disk. Remove from pool.
4. Flush: Write specific dirty page to disk (e.g., for checkpointing).

The strict dichotomy between RAM and disk is beginning to blur as new technologies fill the gap with intermediate performance and persistence characteristics.

Persistent Memory (PMEM):

Intel Optane Persistent Memory (and similar technologies) provides:

• Byte-addressable access: Like RAM, accessed directly by CPU load/store instructions
• Persistence: Data survives power loss like SSD
• Latency: ~300ns—between RAM (~100ns) and SSD (~25μs)
• Capacity: 128GB-512GB per DIMM, larger than RAM, smaller than SSD

Databases can use PMEM as:

• An extension of the buffer pool (larger cache)
• Persistent transaction logs (durability without fsync overhead) -Redo log buffers (smaller durability latency)

Storage Class Memory (SCM):

A category including various technologies:

• 3D XPoint (Optane)
• ReRAM
• Phase-Change Memory
• MRAM

All offer persistence with lower latency than NAND flash.

Computational Storage:

Moving computation closer to storage:

• SSDs with embedded processors
• In-storage database filtering
• Reduces data movement across the hierarchy

The divide between RAM and disk (SSD and HDD) is the single most important performance consideration in database systems. Let's consolidate our understanding:

What's Next:

We've examined the performance differences between RAM and disk. But what determines the actual latency of a storage access? The next page dives into Access Times—dissecting the physics and engineering behind storage latency, from CPU clock cycles to disk rotational delays.